RAPIDS Accelerator for Apache Spark Configuration

RAPIDS Accelerator Deployment Guide (0.1.0)

The following is the list of options that rapids-plugin-4-spark supports.

On startup use: --conf [conf key]=[conf value]. For example:

Copy
Copied!
            

${SPARK_HOME}/bin/spark-shell --jars rapids-4-spark_2.12-23.02.0-cuda11.jar \ --conf spark.plugins=com.nvidia.spark.SQLPlugin \ --conf spark.rapids.sql.concurrentGpuTasks=2

At runtime use: spark.conf.set("[conf key]", [conf value]). For example:

Copy
Copied!
            

scala> spark.conf.set("spark.rapids.sql.concurrentGpuTasks", 2)

All configs can be set on startup, but some configs, especially for shuffle, will not work if they are set at runtime. Please check the column of “Applicable at” to see when the config can be set. “Startup” means only valid on startup, “Runtime” means valid on both startup and runtime.

Name

Description

Default Value

Applicable at

spark.rapi ds.alluxio.aut omount.enabled Enable the feature of auto mounting the cloud storage to Alluxio. It requires the Alluxio master is the same node of Spark driver node. When it’s true, it requires an environment variable ALLUXIO_HOME be set properly. The default value of ALLUXIO_HOME is “/opt/a lluxio-2.8.0”. You can set it as an environment variable when running a spark-submit or you can use spark.ya rn.appMasterEn v.ALLUXIO_HOME to set it on Yarn. The Alluxio master’s host and port will be read from alluxio.m aster.hostname and allu xio.master.rpc .port(default: 19998) from ALLUXIO_HOME/c onf/alluxio-si te.properties, then replace a cloud path which matches spark .rapids.alluxi o.bucket.regex like “s 3://bar/b.csv” to “alluxio ://0.1.2.3:199 98/bar/b.csv”, and the bucket “s3://bar” will be mounted to “/bar” in Alluxio automatically. false Runtime
spark .rapids.alluxi o.bucket.regex A regex to decide which bucket should be auto-mounted to Alluxio. E.g. when setting as “^s3 ://bucket.*”, the bucket which starts with “s3://bucket” will be mounted to Alluxio and the path “s3://buc ket-foo/a.csv” will be replaced to “ alluxio://0.1. 2.3:19998/buck et-foo/a.csv”. It’s only valid when setting sp ark.rapids.all uxio.automount .enabled=true. The default value matches all the buckets in “s3://” or “s3a://” scheme. ^ s3a{0,1}://.* Runtime
spark.rapids. alluxio.large. file.threshold The threshold is used to identify whether average size of files is large when reading from S3. If reading large files from S3 and the disks used by Alluxio are slow, directly reading from S3 is better than reading caches from Alluxio, because S3 network bandwidth is faster than local disk. This improvement takes effect when sp ark.rapids.all uxio.slow.disk is enabled. 67108864 Runtime
spark.r apids.alluxio. pathsToReplace List of paths to be replaced with corresponding Alluxio scheme. E.g. when configure is set to “s3://fo o->alluxio://0 .1.2.3:19998/f oo,gs://bar->a lluxio://0.1.2 .3:19998/bar”, it means: “s 3://foo/a.csv” will be replaced to “alluxi o://0.1.2.3:19 998/foo/a.csv” and “g s://bar/b.csv” will be replaced to “alluxio ://0.1.2.3:199 98/bar/b.csv”. To use this config, you have to mount the buckets to Alluxio by yourself. If you set this config, spark.rapi ds.alluxio.aut omount.enabled won’t be valid. None Startup
spark.rap ids.alluxio.re placement.algo The algorithm used when replacing the UFS path with the Alluxio path. CONVERT_TIME and TASK_TIME are the valid options. CONVERT_TIME indicates that we do it when we convert it to a GPU file read, this has extra overhead of creating an entirely new file index, which requires listing the files and getting all new file info from Alluxio. TASK_TIME replaces the path as late as possible inside of the task. By waiting and replacing it at task time, it just replaces the path without fetching the file information again, this is faster but doesn’t update locality information if that has a bit impact on performance. TASK_TIME Runtime
sp ark.rapids.all uxio.slow.disk Indicates whether the disks used by Alluxio are slow. If it’s true and reading S3 large files, Rapids Accelerator reads from S3 directly instead of reading from Alluxio caches. Refer to spark.rapids. alluxio.large. file.threshold which defines a threshold that identifying whether files are large. Typically, it’s slow disks if speed is less than 300M/second. If using convert time spark.rapi ds.alluxio.rep lacement.algo, this may not apply to all file types like Delta files true Runtime
spark.rapid s.alluxio.user Alluxio user is set on the Alluxio client, which is used to mount or get information. By default it should be the user that running the Alluxio processes. The default value is ubuntu. ubuntu Runtime
spark.rapid s.cloudSchemes Comma separated list of additional URI schemes that are to be considered cloud based filesystems. Schemes already included: abfs, abfss, dbfs, gs, s3, s3a, s3n, wasbs. Cloud based stores generally would be total separate from the executors and likely have a higher I/O read cost. Many times the cloud filesystems also get better throughput when you have multiple readers in parallel. This is used with s park.rapids.sq l.format.parqu et.reader.type None Runtime
s park.rapids.gp u.resourceName The name of the Spark resource that represents a GPU that you want the plugin to use if using custom resources with Spark. gpu Startup
spark.rap ids.memory.gpu .allocFraction The fraction of available (free) GPU memory that should be allocated for pooled memory. This must be less than or equal to the maximum limit configured via spark.rapids. memory.gpu.max AllocFraction, and greater than or equal to the minimum limit configured via spark.rapids. memory.gpu.min AllocFraction. 1.0 Startup
s park.rapids.me mory.gpu.debug Provides a log of GPU memory allocations and frees. If set to STDOUT or STDERR the logging will go there. Setting it to NONE disables logging. All other values are reserved for possible future expansion and in the mean time will disable logging. NONE Startup
spark.rapi ds.memory.gpu. direct.storage .spill.batchWr iteBuffer.size The size of the GPU memory buffer used to batch small buffers when spilling to GDS. Note that this buffer is mapped to the PCI Base Address Register (BAR) space, which may be very limited on some GPUs (e.g. the NVIDIA T4 only has 256 MiB), and it is also used by UCX bounce buffers. 8388608 Startup
spark.rapi ds.memory.gpu. direct.storage .spill.enabled Should GPUDirect Storage (GDS) be used to spill GPU memory buffers directly to disk. GDS must be enabled and the directory spa rk.local.dir must support GDS. This is an experimental feature. For more information on GDS, see h ttps://docs.nv idia.com/gpudi rect-storage/. false Startup
spark.rapids .memory.gpu.ma xAllocFraction The fraction of total GPU memory that limits the maximum size of the RMM pool. The value must be greater than or equal to the setting for spark.rapi ds.memory.gpu. allocFraction. Note that this limit will be reduced by the reserve memory configured in spar k.rapids.memor y.gpu.reserve. 1.0 Startup
spark.rapids .memory.gpu.mi nAllocFraction The fraction of total GPU memory that limits the minimum size of the RMM pool. The value must be less than or equal to the setting for spark.rapi ds.memory.gpu. allocFraction. 0.25 Startup
spark. rapids.memory. gpu.oomDumpDir The path to a local directory where a heap dump will be created if the GPU encounters an unrecoverable out-of-memory (OOM) error. The filename will be of the form: “gp u-oom–.hprof” where is the process ID, and the dumpId is a sequence number to disambiguate multiple heap dumps per process lifecycle None Startup
spark.rapids.m emory.gpu.pool Select the RMM pooling allocator to use. Valid values are “DEFAULT”, “ARENA”, “ASYNC”, and “NONE”. With “DEFAULT”, the RMM pool allocator is used; with “ARENA”, the RMM arena allocator is used; with “ASYNC”, the new CUDA stream-ordered memory allocator in CUDA 11.2+ is used. If set to “NONE”, pooling is disabled and RMM just passes through to CUDA memory allocation directly. ASYNC Startup
spark.rapid s.memory.gpu.p ooling.enabled Should RMM act as a pooling allocator for GPU memory, or should it just pass through to CUDA memory allocation directly. DEPRECATED: please use spark.rapids.m emory.gpu.pool instead. true Startup
spa rk.rapids.memo ry.gpu.reserve The amount of GPU memory that should remain unallocated by RMM and left for system use such as memory needed for kernels and kernel launches. 671088640 Startup
spark.rapid s.memory.gpu.u nspill.enabled When a spilled GPU buffer is needed again, should it be unspilled, or only copied back into GPU memory temporarily. Unspilling may be useful for GPU buffers that are needed frequently, for example, broadcast variables; however, it may also increase GPU memory usage false Startup
spark.rapids.m emory.host.pag eablePool.size The size of the pageable memory pool in bytes unless otherwise specified. Use 0 to disable the pool. 1073741824 Startup
spark.rapids. memory.host.sp illStorageSize Amount of off-heap host memory to use for buffering spilled GPU data before spilling to local disk. Use -1 to set the amount to the combined size of pinned and pageable memory pools. -1 Startup
spark.r apids.memory.p innedPool.size The size of the pinned memory pool in bytes unless otherwise specified. Use 0 to disable the pool. 0 Startup
s park.rapids.py thon.concurren tPythonWorkers Set the number of Python worker processes that can execute concurrently per GPU. Python worker processes may temporarily block when the number of concurrent Python worker processes started by the same executor exceeds this amount. Allowing too many concurrent tasks on the same GPU may lead to GPU out of memory errors. >0 means enabled, while <=0 means unlimited 0 Runtime
sp ark.rapids.pyt hon.memory.gpu .allocFraction The fraction of total GPU memory that should be initially allocated for pooled memory for all the Python workers. It supposes to be less than (1 - $(spark.rapids .memory.gpu.al locFraction)), since the executor will share the GPU with its owning Python workers. Half of the rest will be used if not specified None Runtime
spark .rapids.python .memory.gpu.ma xAllocFraction The fraction of total GPU memory that limits the maximum size of the RMM pool for all the Python workers. It supposes to be less than (1 - $(s park.rapids.me mory.gpu.maxAl locFraction)), since the executor will share the GPU with its owning Python workers. when setting to 0 it means no limit. 0.0 Runtime
spar k.rapids.pytho n.memory.gpu.p ooling.enabled Should RMM in Python workers act as a pooling allocator for GPU memory, or should it just pass through to CUDA memory allocation directly. When not specified, It will honor the value of config ‘spark.rapids .memory.gpu.po oling.enabled’ None Runtime
spark.rapids.s huffle.enabled Enable or disable the RAPIDS Shuffle Manager at runtime. The RAPIDS Shuffle Ma nager must already be configured. When set to false, the built-in Spark shuffle will be used. true Runtime
spark.rapid s.shuffle.mode RAPIDS Shuffle Manager mode. “M ULTITHREADED”: shuffle file writes and reads are parallelized using a thread pool. “UCX”: (requires UCX installation) uses accelerated transports for transferring shuffle blocks. “CACHE_ONLY”: use when running a single executor, for short-circuit cached shuffle (for testing purposes). MULTITHREADED Startup
spark.rap ids.shuffle.mu ltiThreaded.ma xBytesInFlight The size limit, in bytes, that the RAPIDS shuffle manager configured in “ MULTITHREADED” mode will allow to be deserialized concurrently per task. This is also the maximum amount of memory that will be used per task. This should ideally be at least the same size as the batch size so we don’t have to wait to process a single batch. 2147483647 Startup
spark.r apids.shuffle. multiThreaded. reader.threads The number of threads to use for reading shuffle blocks per executor in the RAPIDS shuffle manager configured in “ MULTITHREADED” mode. There are two special values: 0 = feature is disabled, falls back to Spark built-in shuffle reader; 1 = our implementation of Spark’s built-in shuffle reader with extra metrics. 20 Startup
spark.r apids.shuffle. multiThreaded. writer.threads The number of threads to use for writing shuffle blocks per executor in the RAPIDS shuffle manager configured in “ MULTITHREADED” mode. There are two special values: 0 = feature is disabled, falls back to Spark built-in shuffle writer; 1 = our implementation of Spark’s built-in shuffle writer with extra metrics. 20 Startup
spark.rapids. shuffle.transp ort.earlyStart Enable early connection establishment for RAPIDS Shuffle true Startup
spa rk.rapids.shuf fle.transport. earlyStart.hea rtbeatInterval Shuffle early start heartbeat interval ( milliseconds). Executors will send a heartbeat RPC message to the driver at this interval 5000 Startup
sp ark.rapids.shu ffle.transport .earlyStart.he artbeatTimeout Shuffle early start heartbeat timeout ( milliseconds). Executors that don’t heartbeat within this timeout will be considered stale. This timeout must be higher than the value for spa rk.rapids.shuf fle.transport. earlyStart.hea rtbeatInterval 10000 Startup
spark.rapids .shuffle.trans port.maxReceiv eInflightBytes Maximum aggregate amount of bytes that be fetched at any given time from peers during shuffle 1073741824 Startup
spark.r apids.shuffle. ucx.activeMess ages.forceRndv Set to true to force ‘rndv’ mode for all UCX Active Messages. This should only be required with UCX 1.10.x. UCX 1.11.x deployments should set to false. false Startup
spa rk.rapids.shuf fle.ucx.manage mentServerHost The host to be used to start the management server null Startup
spark. rapids.shuffle .ucx.useWakeup When set to true, use UCX’s event-based progress (epoll) in order to wake up the progress thread when needed, instead of a hot loop. true Startup
spa rk.rapids.sql. batchSizeBytes Set the target number of bytes for a GPU batch. Splits sizes for input data is covered by separate configs. The maximum setting is 2 GB to avoid exceeding the cudf row count limit of a column. 2147483647 Runtime
s park.rapids.sq l.castDecimalT oFloat.enabled Casting from decimal to floating point types on the GPU returns results that have tiny difference compared to results returned from CPU. true Runtime
sp ark.rapids.sql .castDecimalTo String.enabled When set to true, casting from decimal to string is supported on the GPU. The GPU does NOT produce exact same string as spark produces, but producing strings which are semantically equal. For instance, given input B igDecimal(123, -2), the GPU produces “12300”, which spark produces “1.23E+4”. false Runtime
s park.rapids.sq l.castFloatToD ecimal.enabled Casting from floating point types to decimal on the GPU returns results that have tiny difference compared to results returned from CPU. true Runtime
spark.r apids.sql.cast FloatToIntegra lTypes.enabled Casting from floating point types to integral types on the GPU supports a slightly different range of values when using Spark 3.1.0 or later. Refer to the CAST documentation for more details. true Runtime
spark.rapids.s ql.castFloatTo String.enabled Casting from floating point types to string on the GPU returns results that have a different precision than the default results of Spark. true Runtime
spark.rapids.s ql.castStringT oFloat.enabled When set to true, enables casting from strings to float types (float, double) on the GPU. Currently hex values aren’t supported on the GPU. Also note that casting from string to float types on the GPU returns incorrect results when the string represents any number “1.7976931 348623158E308” <= x < “1.7976931 348623159E308” and “-1.7976931 348623158E308” >= x > “-1.7976931 348623159E308” in both these cases the GPU returns D ouble.MaxValue while CPU returns “+Infinity” and “-Infinity” respectively true Runtime
spar k.rapids.sql.c astStringToTim estamp.enabled When set to true, casting from string to timestamp is supported on the GPU. The GPU only supports a subset of formats when casting strings to timestamps. Refer to the CAST documentation for more details. false Runtime
spark.rapi ds.sql.coalesc ing.reader.num FilterParallel This controls the number of files the coalescing reader will run in each thread when it filters blocks for reading. If this value is greater than zero the files will be filtered in a multithreaded manner where each thread filters the number of files set by this config. If this is set to zero the files are filtered serially. This uses the same thread pool as the multithreaded reader, see spar k.rapids.sql.m ultiThreadedRe ad.numThreads. Note that filtering multithreaded is useful with Alluxio. 0 Runtime
spark.r apids.sql.conc urrentGpuTasks Set the number of tasks that can execute concurrently per GPU. Tasks may temporarily block when the number of concurrent tasks in the executor exceeds this amount. Allowing too many concurrent tasks on the same GPU may lead to GPU out of memory errors. 1 Runtime
spark.rap ids.sql.concur rentWriterPart itionFlushSize The flush size of the concurrent writer cache in bytes for each partition. If specified spark.sql.max ConcurrentOutp utFileWriters, use concurrent writer to write data. Concurrent writer first caches data for each partition and begins to flush the data if it finds one partition with a size that is greater than or equal to this config. The default value is 0, which will try to select a size based off of file type specific configs. E.g.: It uses write.pa rquet.row-grou p-size-bytes config for Parquet type and orc .stripe.size config for Orc type. If the value is greater than 0, will use this positive value.Max value may get better performance but not always, because concurrent writer uses spillable cache and big value may cause more IO swaps. 0 Runtime
spark.rapids. sql.csv.read.d ecimal.enabled CSV reading is not 100% compatible when reading decimals. false Runtime
spark.rapids .sql.csv.read. double.enabled CSV reading is not 100% compatible when reading doubles. true Runtime
spark.rapid s.sql.csv.read .float.enabled CSV reading is not 100% compatible when reading floats. true Runtime
spark.rapids.s ql.decimalOver flowGuarantees FOR TESTING ONLY. DO NOT USE IN PRODUCTION. Please see the decimal section of the compatibility documents for more information on this config. true Runtime
spa rk.rapids.sql. detectDeltaChe ckpointQueries Queries against Delta Lake _delta_log checkpoint Parquet files are not efficient on the GPU. When this option is enabled, the plugin will attempt to detect these queries and fall back to the CPU. true Runtime
spark.rapi ds.sql.detectD eltaLogQueries Queries against Delta Lake _delta_log JSON files are not efficient on the GPU. When this option is enabled, the plugin will attempt to detect these queries and fall back to the CPU. true Runtime
spark.rapi ds.sql.enabled Enable (true) or disable (false) sql operations on the GPU true Runtime
spark.rapi ds.sql.explain Explain why some parts of a query were not placed on a GPU or not. Possible values are ALL: print everything, NONE: print nothing, NOT_ON_GPU: print only parts of a query that did not go on the GPU NOT_ON_GPU Runtime
spark.rapids.s ql.fast.sample Option to turn on fast sample. If enable it is inconsistent with CPU sample because of GPU sample algorithm is inconsistent with CPU. false Runtime
spark.ra pids.sql.forma t.avro.enabled When set to true enables all avro input and output acceleration. (only input is currently supported anyways) false Runtime
spark.rapi ds.sql.format. avro.multiThre adedRead.maxNu mFilesParallel A limit on the maximum number of files per task processed in parallel on the CPU side before the file is sent to the GPU. This affects the amount of host memory used when reading the files in parallel. Used with MULTITHREADED reader, see spark.rapids. sql.format.avr o.reader.type. 2147483647 Runtime
s park.rapids.sq l.format.avro. multiThreadedR ead.numThreads The maximum number of threads, on one executor, to use for reading small Avro files in parallel. This can not be changed at runtime after the executor has started. Used with MULTITHREADED reader, see spark.rapids. sql.format.avr o.reader.type. DEPRECATED: use spa rk.rapids.sql. multiThreadedR ead.numThreads None Startup
spark.rapids. sql.format.avr o.read.enabled When set to true enables avro input acceleration false Runtime
spark.rapids .sql.format.av ro.reader.type Sets the Avro reader type. We support different types that are optimized for different environments. The original Spark style reader can be selected by setting this to PERFILE which individually reads and copies files to the GPU. Loading many small files individually has high overhead, and using either COALESCING or MULTITHREADED is recommended instead. The COALESCING reader is good when using a local file system where the executors are on the same nodes or close to the nodes the data is being read on. This reader coalesces all the files assigned to a task into a single host buffer before sending it down to the GPU. It copies blocks from a single file into a host buffer in separate threads in parallel, see spar k.rapids.sql.m ultiThreadedRe ad.numThreads. MULTITHREADED is good for cloud environments where you are reading from a blobstore that is totally separate and likely has a higher I/O read cost. Many times the cloud environments also get better throughput when you have multiple readers in parallel. This reader uses multiple threads to read each file in parallel and each file is sent to the GPU separately. This allows the CPU to keep reading while GPU is also doing work. See spa rk.rapids.sql. multiThreadedR ead.numThreads and spark.rapi ds.sql.format. avro.multiThre adedRead.maxNu mFilesParallel to control the number of threads and amount of memory used. By default this is set to AUTO so we select the reader we think is best. This will either be the COALESCING or the MULTITHREADED based on whether we think the file is in the cloud. See spark.rapids .cloudSchemes. AUTO Runtime
spark.r apids.sql.form at.csv.enabled When set to false disables all csv input and output acceleration. (only input is currently supported anyways) true Runtime
spark.rapids .sql.format.cs v.read.enabled When set to false disables csv input acceleration true Runtime
s park.rapids.sq l.format.delta .write.enabled When set to false disables Delta Lake output acceleration. true Runtime
spark.rapids. sql.format.hiv e.text.enabled When set to false disables Hive text table acceleration true Runtime
spark.rapids .sql.format.hi ve.text.read.d ecimal.enabled Hive text file reading is not 100% compatible when reading decimals. Hive has more limitations on what is valid compared to the GPU implementation in some corner cases. See https:// github.com/NVI DIA/spark-rapi ds/issues/7246 true Runtime
spark.rapid s.sql.format.h ive.text.read. double.enabled Hive text file reading is not 100% compatible when reading doubles. true Runtime
spar k.rapids.sql.f ormat.hive.tex t.read.enabled When set to false disables Hive text table read acceleration true Runtime
spark.rapi ds.sql.format. hive.text.read .float.enabled Hive text file reading is not 100% compatible when reading floats. true Runtime
spark .rapids.sql.fo rmat.hive.text .write.enabled When set to false disables Hive text table write acceleration false Runtime
spark.rapid s.sql.format.i ceberg.enabled When set to false disables all Iceberg acceleration true Runtime
sp ark.rapids.sql .format.iceber g.read.enabled When set to false disables Iceberg input acceleration true Runtime
spark.ra pids.sql.forma t.json.enabled When set to true enables all json input and output acceleration. (only input is currently supported anyways) false Runtime
spark.rapids. sql.format.jso n.read.enabled When set to true enables json input acceleration false Runtime
spark.r apids.sql.form at.orc.enabled When set to false disables all orc input and output acceleration true Runtime
spark.rapid s.sql.format.o rc.floatTypesT oString.enable When reading an ORC file, the source data s chemas(schemas of ORC file) may differ from the target schemas (schemas of the reader), we need to handle the castings from source type to target type. Since float/double numbers in GPU have different precision with CPU, when casting float/double to string, the result of GPU is different from result of CPU spark. Its default value is true (this means the strings result will differ from result of CPU). If it’s set false explicitly and there exists casting from float/double to string in the job, then such behavior will cause an exception, and the job will fail. true Runtime
spark.rap ids.sql.format .orc.multiThre adedRead.maxNu mFilesParallel A limit on the maximum number of files per task processed in parallel on the CPU side before the file is sent to the GPU. This affects the amount of host memory used when reading the files in parallel. Used with MULTITHREADED reader, see spark.rapids .sql.format.or c.reader.type. 2147483647 Runtime
spark.rapids.s ql.format.orc. multiThreadedR ead.numThreads The maximum number of threads, on the executor, to use for reading small ORC files in parallel. This can not be changed at runtime after the executor has started. Used with MULTITHREADED reader, see spark.rapids .sql.format.or c.reader.type. DEPRECATED: use spa rk.rapids.sql. multiThreadedR ead.numThreads None Startup
spark.rapids .sql.format.or c.read.enabled When set to false disables orc input acceleration true Runtime
spark.rapid s.sql.format.o rc.reader.type Sets the ORC reader type. We support different types that are optimized for different environments. The original Spark style reader can be selected by setting this to PERFILE which individually reads and copies files to the GPU. Loading many small files individually has high overhead, and using either COALESCING or MULTITHREADED is recommended instead. The COALESCING reader is good when using a local file system where the executors are on the same nodes or close to the nodes the data is being read on. This reader coalesces all the files assigned to a task into a single host buffer before sending it down to the GPU. It copies blocks from a single file into a host buffer in separate threads in parallel, see spar k.rapids.sql.m ultiThreadedRe ad.numThreads. MULTITHREADED is good for cloud environments where you are reading from a blobstore that is totally separate and likely has a higher I/O read cost. Many times the cloud environments also get better throughput when you have multiple readers in parallel. This reader uses multiple threads to read each file in parallel and each file is sent to the GPU separately. This allows the CPU to keep reading while GPU is also doing work. See spa rk.rapids.sql. multiThreadedR ead.numThreads and spark.rap ids.sql.format .orc.multiThre adedRead.maxNu mFilesParallel to control the number of threads and amount of memory used. By default this is set to AUTO so we select the reader we think is best. This will either be the COALESCING or the MULTITHREADED based on whether we think the file is in the cloud. See spark.rapids .cloudSchemes. AUTO Runtime
spark.rapids. sql.format.orc .write.enabled When set to false disables orc output acceleration true Runtime
spark.rapid s.sql.format.p arquet.enabled When set to false disables all parquet input and output acceleration true Runtime
spark.rapids. sql.format.par quet.multiThre adedRead.maxNu mFilesParallel A limit on the maximum number of files per task processed in parallel on the CPU side before the file is sent to the GPU. This affects the amount of host memory used when reading the files in parallel. Used with MULTITHREADED reader, see sp ark.rapids.sql .format.parque t.reader.type. 2147483647 Runtime
spar k.rapids.sql.f ormat.parquet. multiThreadedR ead.numThreads The maximum number of threads, on the executor, to use for reading small Parquet files in parallel. This can not be changed at runtime after the executor has started. Used with COALESCING and MULTITHREADED reader, see sp ark.rapids.sql .format.parque t.reader.type. DEPRECATED: use spa rk.rapids.sql. multiThreadedR ead.numThreads None Startup
spark.r apids.sql.form at.parquet.mul tithreaded.com bine.sizeBytes The target size in bytes to combine multiple small files together when using the MULTITHREADED parquet reader. With combine disabled, the MULTITHREADED reader reads the files in parallel and sends individual files down to the GPU, but that can be inefficient for small files. When combine is enabled, files that are ready within spark. rapids.sql.for mat.parquet.mu ltithreaded.co mbine.waitTime together, up to this threshold size, are combined before sending down to GPU. This can be disabled by setting it to 0. Note that combine also will not go over the spark.rap ids.sql.reader .batchSizeRows or spark.rapi ds.sql.reader. batchSizeBytes limits. 67108864 Runtime
spark. rapids.sql.for mat.parquet.mu ltithreaded.co mbine.waitTime When using the multithreaded parquet reader with combine mode, how long to wait, in milliseconds, for more files to finish if haven’t met the size threshold. Note that this will wait this amount of time from when the last file was available, so total wait time could be larger then this. 200 Runtime
spar k.rapids.sql.f ormat.parquet. multithreaded. read.keepOrder When using the MULTITHREADED reader, if this is set to true we read the files in the same order Spark does, otherwise the order may not be the same. true Runtime
sp ark.rapids.sql .format.parque t.read.enabled When set to false disables parquet input acceleration true Runtime
spark.ra pids.sql.forma t.parquet.read er.footer.type In some cases reading the footer of the file is very expensive. Typically this happens when there are a large number of columns and relatively few of them are being read on a large number of files. This provides the ability to use a different path to parse and filter the footer. AUTO is the default and decides which path to take using a heuristic. JAVA follows closely with what Apache Spark does. NATIVE will parse and filter the footer using C++. AUTO Runtime
s park.rapids.sq l.format.parqu et.reader.type Sets the Parquet reader type. We support different types that are optimized for different environments. The original Spark style reader can be selected by setting this to PERFILE which individually reads and copies files to the GPU. Loading many small files individually has high overhead, and using either COALESCING or MULTITHREADED is recommended instead. The COALESCING reader is good when using a local file system where the executors are on the same nodes or close to the nodes the data is being read on. This reader coalesces all the files assigned to a task into a single host buffer before sending it down to the GPU. It copies blocks from a single file into a host buffer in separate threads in parallel, see spar k.rapids.sql.m ultiThreadedRe ad.numThreads. MULTITHREADED is good for cloud environments where you are reading from a blobstore that is totally separate and likely has a higher I/O read cost. Many times the cloud environments also get better throughput when you have multiple readers in parallel. This reader uses multiple threads to read each file in parallel and each file is sent to the GPU separately. This allows the CPU to keep reading while GPU is also doing work. See spa rk.rapids.sql. multiThreadedR ead.numThreads and spark.rapids. sql.format.par quet.multiThre adedRead.maxNu mFilesParallel to control the number of threads and amount of memory used. By default this is set to AUTO so we select the reader we think is best. This will either be the COALESCING or the MULTITHREADED based on whether we think the file is in the cloud. See spark.rapids .cloudSchemes. AUTO Runtime
spa rk.rapids.sql. format.parquet .write.enabled When set to false disables parquet output acceleration true Runtime
spark.rapi ds.sql.format. parquet.writer .int96.enabled When set to false, disables accelerated parquet write if the spark.sql .parquet.outpu tTimestampType is set to INT96 true Runtime
spark.rapi ds.sql.hasExte ndedYearValues Spark 3.2.0+ extended parsing of years in dates and timestamps to support the full range of possible values. Prior to this it was limited to a positive 4 digit year. The Accelerator does not support the extended range yet. This config indicates if your data includes this extended range or not, or if you don’t care about getting the correct values on values with the extended range. true Runtime
spark.rapids. sql.hashOptimi zeSort.enabled Whether sorts should be inserted after some hashed operations to improve output ordering. This can improve output file sizes when saving to columnar formats. false Runtime
spark.rapids. sql.improvedFl oatOps.enabled For some floating point operations spark uses one way to compute the value and the underlying cudf implementation can use an improved algorithm. In some cases this can result in cudf producing an answer when spark overflows. true Runtime
spark.rapids .sql.improvedT imeOps.enabled When set to true, some operators will avoid overflowing by converting epoch days directly to seconds without first converting to microseconds false Runtime
spark. rapids.sql.inc ompatibleDateF ormats.enabled When parsing strings as dates and timestamps in functions like u nix_timestamp, some formats are fully supported on the GPU and some are unsupported and will fall back to the CPU. Some formats behave differently on the GPU than the CPU. Spark on the CPU interprets date formats with unsupported trailing characters as nulls, while Spark on the GPU will parse the date with invalid trailing characters. More detail can be found at parsing strings as dates or times tamps. false Runtime
spark.rapids .sql.incompati bleOps.enabled For operations that work, but are not 100% compatible with the Spark equivalent set if they should be enabled by default or disabled by default. true Runtime
spark.r apids.sql.join .cross.enabled When set to true cross joins are enabled on the GPU true Runtime
spark.rapid s.sql.join.exi stence.enabled When set to true existence joins are enabled on the GPU true Runtime
spark.rapid s.sql.join.ful lOuter.enabled When set to true full outer joins are enabled on the GPU true Runtime
spark.r apids.sql.join .inner.enabled When set to true inner joins are enabled on the GPU true Runtime
spark.rapi ds.sql.join.le ftAnti.enabled When set to true left anti joins are enabled on the GPU true Runtime
spark.rapid s.sql.join.lef tOuter.enabled When set to true left outer joins are enabled on the GPU true Runtime
spark.rapi ds.sql.join.le ftSemi.enabled When set to true left semi joins are enabled on the GPU true Runtime
spark.rapids .sql.join.righ tOuter.enabled When set to true right outer joins are enabled on the GPU true Runtime
spark.rapids.s ql.json.read.d ecimal.enabled JSON reading is not 100% compatible when reading decimals. false Runtime
spark.rapids. sql.json.read. double.enabled JSON reading is not 100% compatible when reading doubles. true Runtime
spark.rapids .sql.json.read .float.enabled JSON reading is not 100% compatible when reading floats. true Runtime
sp ark.rapids.sql .metrics.level GPU plans can produce a lot more metrics than CPU plans do. In very large queries this can sometimes result in going over the max result size limit for the driver. Supported values include DEBUG which will enable all metrics supported and typically only needs to be enabled when debugging the plugin. MODERATE which should output enough metrics to understand how long each part of the query is taking and how much data is going to each part of the query. ESSENTIAL which disables most metrics except those Apache Spark CPU plans will also report or their equivalents. MODERATE Runtime
spark.r apids.sql.mode Set the mode for the Rapids Accelerator. The supported modes are explainOnly and executeOnGPU. This config can not be changed at runtime, you must restart the application for it to take affect. The default mode is executeOnGPU, which means the RAPIDS Accelerator plugin convert the Spark operations and execute them on the GPU when possible. The explainOnly mode allows running queries on the CPU and the RAPIDS Accelerator will evaluate the queries as if it was going to run on the GPU. The explanations of what would have run on the GPU and why are output in log messages. When using explainOnly mode, the default explain output is ALL, this can be changed by setting spark.rapid s.sql.explain. See that config for more details. executeongpu Startup
spa rk.rapids.sql. multiThreadedR ead.numThreads The maximum number of threads on each executor to use for reading small files in parallel. This can not be changed at runtime after the executor has started. Used with COALESCING and MULTITHREADED readers, see sp ark.rapids.sql .format.parque t.reader.type, spark.rapids .sql.format.or c.reader.type, or spark.rapids .sql.format.av ro.reader.type for a discussion of reader types. If it is not set explicitly and spark. executor.cores is set, it will be tried to assign value of max(MULTIT HREAD_READ_NUM _THREADS_DEFAU LT, spark.exec utor.cores), where MULTITHR EAD_READ_NUM_T HREADS_DEFAULT = 20. 20 Startup
spark.r apids.sql.pyth on.gpu.enabled This is an experimental feature and is likely to change in the future. Enable (true) or disable (false) support for scheduling Python Pandas UDFs with GPU resources. When enabled, pandas UDFs are assumed to share the same GPU that the RAPIDs accelerator uses and will honor the python GPU configs false Runtime
spark.rapi ds.sql.reader. batchSizeBytes Soft limit on the maximum number of bytes the reader reads per batch. The readers will read chunks of data until this limit is met or exceeded. Note that the reader may estimate the number of bytes that will be used on the GPU in some cases based on the schema and number of rows in each batch. 2147483647 Runtime
spark.rap ids.sql.reader .batchSizeRows Soft limit on the maximum number of rows the reader will read per batch. The orc and parquet readers will read row groups until this limit is met or exceeded. The limit is respected by the csv reader. 2147483647 Runtime
spa rk.rapids.sql. reader.chunked Enable a chunked reader where possible. A chunked reader allows reading highly compressed data that could not be read otherwise, but at the expense of more GPU memory, and in some cases more GPU computation. true Runtime
spa rk.rapids.sql. regexp.enabled Specifies whether supported regular expressions will be evaluated on the GPU. Unsupported expressions will fall back to CPU. However, there are some known edge cases that will still execute on GPU and produce incorrect results and these are documented in the compatibility guide. Setting this config to false will make all regular expressions run on the CPU instead. true Runtime
s park.rapids.sq l.regexp.maxSt ateMemoryBytes Specifies the maximum memory on GPU to be used for regular e xpressions.The memory usage is an estimate based on an upper-bound approximation on the complexity of the regular expression. Note that the actual memory usage may still be higher than this estimate depending on the number of rows in the datacolumn and the input strings themselves. It is recommended to not set this to more than 3 times spa rk.rapids.sql. batchSizeBytes 2147483647 Runtime
spa rk.rapids.sql. replaceSortMer geJoin.enabled Allow replacing sortMergeJoin with HashJoin true Runtime
spark.ra pids.sql.rowBa sedUDF.enabled When set to true, optimizes a row-based UDF in a GPU operation by transferring only the data it needs between GPU and CPU inside a query operation, instead of falling this operation back to CPU. This is an experimental feature, and this config might be removed in the future. false Runtime
spark.rap ids.sql.shuffl e.spillThreads Number of threads used to spill shuffle data to disk in the background. 6 Runtime
spark.r apids.sql.stab leSort.enabled Enable or disable stable sorting. Apache Spark’s sorting is typically a stable sort, but sort stability cannot be guaranteed in distributed work loads because the order in which upstream data arrives to a task is not guaranteed. Sort stability then only matters when reading and sorting data from a file using a single t ask/partition. Because of limitations in the plugin when you enable stable sorting all of the data for a single task will be combined into a single batch before sorting. This currently disables spilling from GPU memory if the data size is too large. false Runtime
spark.rapids .sql.suppressP lanningFailure Option to fallback an individual query to CPU if an unexpected condition prevents the query plan from being converted to a GPU-enabled one. Note this is different from a normal CPU fallback for a yet-t o-be-supported Spark SQL feature. If this happens the error should be reported and investigated as a GitHub issue. false Runtime
spark.ra pids.sql.udfCo mpiler.enabled When set to true, Scala UDFs will be considered for compilation as Catalyst expressions false Runtime
spark.rapids. sql.variableFl oatAgg.enabled Spark assumes that all operations produce the exact same result each time. This is not true for some floating point aggregations, which can produce slightly different results on the GPU as the aggregation is done in parallel. This can enable those operations if you know the query is only computing it once. true Runtime
spark.rapids.s ql.window.rang e.byte.enabled When the order-by column of a range based window is byte type and the range boundary calculated for a value has overflow, CPU and GPU will get the different results. When set to false disables the range window acceleration for the byte type order-by column false Runtime
spa rk.rapids.sql. window.range.d ecimal.enabled When set to false, this disables the range window acceleration for the DECIMAL type order-by column true Runtime
spark.rapids. sql.window.ran ge.int.enabled When the order-by column of a range based window is int type and the range boundary calculated for a value has overflow, CPU and GPU will get the different results. When set to false disables the range window acceleration for the int type order-by column true Runtime
spark.rapids.s ql.window.rang e.long.enabled When the order-by column of a range based window is long type and the range boundary calculated for a value has overflow, CPU and GPU will get the different results. When set to false disables the range window acceleration for the long type order-by column true Runtime
s park.rapids.sq l.window.range .short.enabled When the order-by column of a range based window is short type and the range boundary calculated for a value has overflow, CPU and GPU will get the different results. When set to false disables the range window acceleration for the short type order-by column false Runtime

The RAPIDS Accelerator for Apache Spark can be configured to enable or disable specific GPU accelerated expressions. Enabled expressions are candidates for GPU execution. If the expression is configured as disabled, the accelerator plugin will not attempt replacement, and it will run on the CPU.

Please leverage the spark.rapids.sql.explain setting to get feedback from the plugin as to why parts of a query may not be executing on the GPU.

Note

Setting spark.rapids.sql.incompatibleOps.enabled=true will enable all the settings in the table below which are not enabled by default due to incompatibilities.

Expressions

Name

SQL Function(s)

Description

Default Value

Notes

spark.rap ids.sql.exp ression.Abs abs Absolute value true None
spark.rapi ds.sql.expr ession.Acos acos Inverse cosine true None
spark.rapid s.sql.expre ssion.Acosh acosh Inverse hyperbolic cosine true None
spark.rap ids.sql.exp ression.Add + Addition true None
spark.rapid s.sql.expre ssion.Alias Gives a column a name true None
spark.rap ids.sql.exp ression.And and Logical AND true None
spa rk.rapids.s ql.expressi on.AnsiCast Convert a column of one type of data into another type true None
spark.ra pids.sql.ex pression.Ar rayContains array _contains Returns a boolean if the array contains the passed in key true None
spark. rapids.sql. expression. ArrayExcept arr ay_except Returns an array of the elements in array1 but not in array2, without duplicates true This is not 100% compatible with the Spark version because the GPU imp lementation treats -0.0 and 0.0 as equal, but the CPU imp lementation currently does not (see SP ARK-39845). Also, Apache Spark 3.1.3 fixed issue SPARK-36741 where NaNs in these set like operators were not treated as being equal. We have chosen to break with co mpatibility for the older versions of Spark in this instance and handle NaNs the same as 3.1.3+
spark. rapids.sql. expression. ArrayExists exists Return true if any element satisfies the predicate Lam bdaFunction true None
spark.rap ids.sql.exp ression.Arr ayIntersect array_ intersect Returns an array of the elements in the i ntersection of array1 and array2, without duplicates true This is not 100% compatible with the Spark version because the GPU imp lementation treats -0.0 and 0.0 as equal, but the CPU imp lementation currently does not (see SP ARK-39845). Also, Apache Spark 3.1.3 fixed issue SPARK-36741 where NaNs in these set like operators were not treated as being equal. We have chosen to break with co mpatibility for the older versions of Spark in this instance and handle NaNs the same as 3.1.3+
spa rk.rapids.s ql.expressi on.ArrayMax `` array_max`` Returns the maximum value in the array true None
spa rk.rapids.s ql.expressi on.ArrayMin `` array_min`` Returns the minimum value in the array true None
spark. rapids.sql. expression. ArrayRemove arr ay_remove Returns the array after removing all elements that equal to the input element (right) from the input array (left) true None
spark. rapids.sql. expression. ArrayRepeat arr ay_repeat Returns the array containing the given input value (left) count (right) times true None
spark.rap ids.sql.exp ression.Arr ayTransform `` transform`` Transform elements in an array using the transform function. This is similar to a map in functional programming true None
spark .rapids.sql .expression .ArrayUnion ar ray_union Returns an array of the elements in the union of array1 and array2, without duplicates. true This is not 100% compatible with the Spark version because the GPU imp lementation treats -0.0 and 0.0 as equal, but the CPU imp lementation currently does not (see SP ARK-39845). Also, Apache Spark 3.1.3 fixed issue SPARK-36741 where NaNs in these set like operators were not treated as being equal. We have chosen to break with co mpatibility for the older versions of Spark in this instance and handle NaNs the same as 3.1.3+
spark.ra pids.sql.ex pression.Ar raysOverlap array s_overlap Returns true if a1 contains at least a non-null element present also in a2. If the arrays have no common element and they are both non-empty and either of them contains a null element null is returned, false otherwise. true This is not 100% compatible with the Spark version because the GPU imp lementation treats -0.0 and 0.0 as equal, but the CPU imp lementation currently does not (see SP ARK-39845). Also, Apache Spark 3.1.3 fixed issue SPARK-36741 where NaNs in these set like operators were not treated as being equal. We have chosen to break with co mpatibility for the older versions of Spark in this instance and handle NaNs the same as 3.1.3+
spar k.rapids.sq l.expressio n.ArraysZip a rrays_zip Returns a merged array of structs in which the N-th struct contains all N-th values of input arrays. true None
spark.rapi ds.sql.expr ession.Asin asin Inverse sine true None
spark.rapid s.sql.expre ssion.Asinh asinh Inverse hyperbolic sine true None
spark.rapid s.sql.expre ssion.AtLea stNNonNulls Checks if number of non null/Nan values is greater than a given value true None
spark.rapi ds.sql.expr ession.Atan atan Inverse tangent true None
spark.rapid s.sql.expre ssion.Atanh atanh Inverse hyperbolic tangent true None
sp ark.rapids. sql.express ion.Attribu teReference References an input column true None
s park.rapids .sql.expres sion.BRound bround Round an expression to d decimal places using HALF_EVEN rounding mode true None
spar k.rapids.sq l.expressio n.BitLength b it_length The bit length of string data true None
spark .rapids.sql .expression .BitwiseAnd & Returns the bitwise AND of the operands true None
spark .rapids.sql .expression .BitwiseNot ~ Returns the bitwise NOT of the operands true None
spar k.rapids.sq l.expressio n.BitwiseOr \| Returns the bitwise OR of the operands true None
spark .rapids.sql .expression .BitwiseXor ^ Returns the bitwise XOR of the operands true None
spa rk.rapids.s ql.expressi on.CaseWhen when CASE WHEN expression true None
spark.rapi ds.sql.expr ession.Cast t imestamp, ` tinyint`, binary, float, `` smallint``, string, ` decimal`, double, ` boolean`, cast, date, int, bigint Convert a column of one type of data into another type true None
spark.rapi ds.sql.expr ession.Cbrt cbrt Cube root true None
spark.rapi ds.sql.expr ession.Ceil ` ceiling`, ceil Ceiling of a number true None
spark.ra pids.sql.ex pression.Ch eckOverflow Ch eckOverflow after arithmetic operations between DecimalType data true None
spa rk.rapids.s ql.expressi on.Coalesce ` coalesce` Returns the first non-null argument if exists. Otherwise, null true None
s park.rapids .sql.expres sion.Concat concat List/String concatenate true None
spa rk.rapids.s ql.expressi on.ConcatWs `` concat_ws`` C oncatenates multiple input strings or array of strings into a single string using a given separator true None
spa rk.rapids.s ql.expressi on.Contains Contains true None
spark.rap ids.sql.exp ression.Cos cos Cosine true None
spark.rapi ds.sql.expr ession.Cosh cosh Hyperbolic cosine true None
spark.rap ids.sql.exp ression.Cot cot Cotangent true None
spark. rapids.sql. expression. CreateArray array Returns an array with the given elements true None
spar k.rapids.sq l.expressio n.CreateMap map Create a map true None
s park.rapids .sql.expres sion.Create NamedStruct name d_struct, struct Creates a struct with the given field names and values true None
spark. rapids.sql. expression. CurrentRow$ Special boundary for a window frame, indicating stopping at the current row true None
sp ark.rapids. sql.express ion.DateAdd ` date_add` Returns the date that is num_days after start_date true None
spark.rapi ds.sql.expr ession.Date AddInterval Adds interval to date true None
spa rk.rapids.s ql.expressi on.DateDiff ` datediff` Returns the number of days from startDate to endDate true None
spark.rapi ds.sql.expr ession.Date FormatClass da te_format Converts timestamp to a value of string in the format specified by the date format true None
sp ark.rapids. sql.express ion.DateSub ` date_sub` Returns the date that is num_days before start_date true None
spark .rapids.sql .expression .DayOfMonth da yofmonth, day Returns the day of the month from a date or timestamp true None
spar k.rapids.sq l.expressio n.DayOfWeek `` dayofweek`` Returns the day of the week (1 = Sunday… 7=Saturday) true None
spar k.rapids.sq l.expressio n.DayOfYear `` dayofyear`` Returns the day of the year from a date or timestamp true None
spar k.rapids.sq l.expressio n.DenseRank d ense_rank Window function that returns the dense rank value within the aggregation window true None
s park.rapids .sql.expres sion.Divide / Division true None
spar k.rapids.sq l.expressio n.ElementAt e lement_at Returns element of array at giv en(1-based) index in value if column is array. Returns value for the given key in value if column is map. true None
spa rk.rapids.s ql.expressi on.EndsWith Ends with true None
spark.ra pids.sql.ex pression.Eq ualNullSafe <=> Check if the values are equal including nulls <=> true None
sp ark.rapids. sql.express ion.EqualTo =, == Check if the values are equal true None
spark.rap ids.sql.exp ression.Exp exp Euler’s number e raised to a power true None
sp ark.rapids. sql.express ion.Explode ` explode`, expl ode_outer Given an input array produces a sequence of rows for each value in the array true None
spark.rapid s.sql.expre ssion.Expm1 expm1 Euler’s number e raised to a power minus 1 true None
spark.rapid s.sql.expre ssion.Floor floor Floor of a number true None
spark.rapid s.sql.expre ssion.FromU TCTimestamp from_utc_ timestamp Render the input UTC timestamp in the input timezone true None
spark.r apids.sql.e xpression.F romUnixTime from _unixtime Get the string from a unix timestamp true None
spark.r apids.sql.e xpression.G etArrayItem Gets the field at ordinal in the Array true None
spar k.rapids.sq l.expressio n.GetArrayS tructFields Extracts the o rdinal-th fields of all array elements for the data with the type of array of struct true None
spark.ra pids.sql.ex pression.Ge tJsonObject get_js on_object Extracts a json object from path true None
spark. rapids.sql. expression. GetMapValue Gets Value from a Map based on a key true None
spark.rap ids.sql.exp ression.Get StructField Gets the named field of the struct true None
spark.r apids.sql.e xpression.G etTimestamp Gets timestamps from strings using given pattern. true None
spark. rapids.sql. expression. GreaterThan > > operator true None
sp ark.rapids. sql.express ion.Greater ThanOrEqual >= >= operator true None
spa rk.rapids.s ql.expressi on.Greatest ` greatest` Returns the greatest value of all parameters, skipping null values true None
spark.rapi ds.sql.expr ession.Hour hour Returns the hour component of the strin g/timestamp true None
spark.rapid s.sql.expre ssion.Hypot hypot Pythagorean addition ( Hypotenuse) of real numbers true None
spark.ra pids.sql.ex pression.If if IF expression true None
spark.ra pids.sql.ex pression.In in IN operator true None
spark.rapid s.sql.expre ssion.InSet INSET operator true None
sp ark.rapids. sql.express ion.InitCap initcap Returns str with the first letter of each word in uppercase. All other letters are in lowercase true This is not 100% compatible with the Spark version because the Unicode version used by cuDF and the JVM may differ, resulting in some corner-case characters not changing case correctly.
spar k.rapids.sq l.expressio n.InputFile BlockLength inp ut_file_blo ck_length Returns the length of the block being read, or -1 if not available true None
spa rk.rapids.s ql.expressi on.InputFil eBlockStart in put_file_bl ock_start Returns the start offset of the block being read, or -1 if not available true None
spark.ra pids.sql.ex pression.In putFileName input_ file_name Returns the name of the file being read, or empty string if not available true None
spark.rap ids.sql.exp ression.Int egralDivide div Division with a integer result true None
spark.rapid s.sql.expre ssion.IsNaN isnan Checks if a value is NaN true None
spar k.rapids.sq l.expressio n.IsNotNull `` isnotnull`` Checks if a value is not null true None
s park.rapids .sql.expres sion.IsNull isnull Checks if a value is null true None
spark.ra pids.sql.ex pression.Js onToStructs `` from_json`` Returns a struct value with the given jsonStr and schema false This is disabled by default because parsing JSON from a column has a large number of issues and should be considered beta quality right now.
spar k.rapids.sq l.expressio n.JsonTuple j son_tuple Returns a tuple like the function get_j son_object, but it takes multiple names. All the input parameters and output column types are string. true None
s park.rapids .sql.expres sion.KnownF loatingPoin tNormalized Tag to prevent redundant no rmalization true None
spark.r apids.sql.e xpression.K nownNotNull Tag an expression as known to not be null true None
spark.rap ids.sql.exp ression.Lag lag Window function that returns N entries behind this one true None
spark.rap ids.sql.exp ression.Lam bdaFunction Holds a higher order SQL function true None
sp ark.rapids. sql.express ion.LastDay ` last_day` Returns the last day of the month which the date belongs to true None
spark.rapi ds.sql.expr ession.Lead lead Window function that returns N entries ahead of this one true None
spark.rapid s.sql.expre ssion.Least least Returns the least value of all parameters, skipping null values true None
s park.rapids .sql.expres sion.Length length, characte r_length, ch ar_length String character length or binary byte length true None
spa rk.rapids.s ql.expressi on.LessThan < < operator true None
spark.rapi ds.sql.expr ession.Less ThanOrEqual <= <= operator true None
spark.rapi ds.sql.expr ession.Like like Like true None
sp ark.rapids. sql.express ion.Literal Holds a static value from the query true None
spark.rap ids.sql.exp ression.Log ln Natural log true None
spark.rapid s.sql.expre ssion.Log10 log10 Log base 10 true None
spark.rapid s.sql.expre ssion.Log1p log1p Natural log 1 + expr true None
spark.rapi ds.sql.expr ession.Log2 log2 Log base 2 true None
spar k.rapids.sq l.expressio n.Logarithm log Log variable base true None
spark.rapid s.sql.expre ssion.Lower lower, lcase String lowercase operator true This is not 100% compatible with the Spark version because the Unicode version used by cuDF and the JVM may differ, resulting in some corner-case characters not changing case correctly.
spark. rapids.sql. expression. MakeDecimal Create a Decimal from an unscaled long value for some aggregation op timizations true None
spar k.rapids.sq l.expressio n.MapConcat m ap_concat Returns the union of all the given maps true None
spark .rapids.sql .expression .MapEntries ma p_entries Returns an unordered array of all entries in the given map true None
spar k.rapids.sq l.expressio n.MapFilter m ap_filter Filters entries in a map using the function true None
sp ark.rapids. sql.express ion.MapKeys ` map_keys` Returns an unordered array containing the keys of the map true None
spar k.rapids.sq l.expressio n.MapValues m ap_values Returns an unordered array containing the values of the map true None
spark.rap ids.sql.exp ression.Md5 md5 MD5 hash operator true None
s park.rapids .sql.expres sion.Minute minute Returns the minute component of the strin g/timestamp true None
spark.rap ids.sql.exp ression.Mon otonicallyI ncreasingID monoton ically_incr easing_id Returns mo notonically increasing 64-bit integers true None
spark.rapid s.sql.expre ssion.Month month Returns the month from a date or timestamp true None
spa rk.rapids.s ql.expressi on.Multiply * Mul tiplication true None
spark. rapids.sql. expression. Murmur3Hash hash Murmur3 hash operator true None
spark.rapid s.sql.expre ssion.NaNvl nanvl Evaluates to left iff left is not NaN, right otherwise true None
spa rk.rapids.s ql.expressi on.NamedLam bdaVariable A parameter to a higher order SQL function true None
spark.rap ids.sql.exp ression.Not !, not Boolean not operator true None
spa rk.rapids.s ql.expressi on.NthValue `` nth_value`` nth window operator true None
spark. rapids.sql. expression. OctetLength oct et_length The byte length of string data true None
spark.ra pids.sql.ex pression.Or or Logical OR true None
spark. rapids.sql. expression. PercentRank per cent_rank Window function that returns the percent rank value within the aggregation window true None
spark.rapi ds.sql.expr ession.Pmod pmod Pmod true None
spark .rapids.sql .expression .PosExplode posexplo de_outer, p osexplode Given an input array produces a sequence of rows for each value in the array true None
spark.rap ids.sql.exp ression.Pow pow, power lhs ^ rhs true None
spark.rapi ds.sql.expr ession.Prec iseTimestam pConversion Expression used internally to convert the Ti mestampType to Long and back without losing precision, i.e. in mi croseconds. Used in time windowing true None
spark.rapid s.sql.expre ssion.Promo tePrecision Promo tePrecision before arithmetic operations between DecimalType data true None
spar k.rapids.sq l.expressio n.PythonUDF UDF run in an external python process. Does not actually run on the GPU, but the transfer of data to/from it can be accelerated true None
sp ark.rapids. sql.express ion.Quarter quarter Returns the quarter of the year for date, in the range 1 to 4 true None
spark.rapid s.sql.expre ssion.RLike rlike Regular expression version of Like true None
spark .rapids.sql .expression .RaiseError ra ise_error Throw an exception true None
spark.rapi ds.sql.expr ession.Rand random, rand Generate a random column with i.i.d. uniformly distributed values in [0, 1) true None
spark.rapi ds.sql.expr ession.Rank rank Window function that returns the rank value within the aggregation window true None
spark.ra pids.sql.ex pression.Re gExpExtract regex p_extract Extract a specific group identified by a regular expression true None
spark.rapid s.sql.expre ssion.RegEx pExtractAll regexp_ex tract_all Extract all strings matching a regular expression co rresponding to the regex group index true None
spark.ra pids.sql.ex pression.Re gExpReplace regex p_replace String replace using a regular expression pattern true None
spar k.rapids.sq l.expressio n.Remainder %, mod Remainder or modulo true None
spark.ra pids.sql.ex pression.Re plicateRows Given an input row replicates the row N times true None
sp ark.rapids. sql.express ion.Reverse reverse Returns a reversed string or an array with reverse order of elements true None
spark.rapi ds.sql.expr ession.Rint rint Rounds up a double value to the nearest double equal to an integer true None
spark.rapid s.sql.expre ssion.Round round Round an expression to d decimal places using HALF_UP rounding mode true None
spar k.rapids.sq l.expressio n.RowNumber r ow_number Window function that returns the index for the row within the aggregation window true None
spa rk.rapids.s ql.expressi on.ScalaUDF User Defined Function, the UDF can choose to implement a RAPIDS accelerated interface to get better p erformance. true None
s park.rapids .sql.expres sion.Second second Returns the second component of the strin g/timestamp true None
spa rk.rapids.s ql.expressi on.Sequence ` sequence` Sequence true None
spar k.rapids.sq l.expressio n.ShiftLeft `` shiftleft`` Bitwise shift left («) true None
spark .rapids.sql .expression .ShiftRight s hiftright Bitwise shift right (») true None
sp ark.rapids. sql.express ion.ShiftRi ghtUnsigned shiftrigh tunsigned Bitwise unsigned shift right (»>) true None
s park.rapids .sql.expres sion.Signum sign, signum Returns -1.0, 0.0 or 1.0 as expr is negative, 0 or positive true None
spark.rap ids.sql.exp ression.Sin sin Sine true None
spark.rapi ds.sql.expr ession.Sinh sinh Hyperbolic sine true None
spark.rapi ds.sql.expr ession.Size size, ca rdinality The size of an array or a map true None
spar k.rapids.sq l.expressio n.SortArray s ort_array Returns a sorted array with the input array and the ascending / descending order true None
spar k.rapids.sq l.expressio n.SortOrder Sort order true None
spark.rapid s.sql.expre ssion.Spark PartitionID spark_par tition_id Returns the current partition id true None
spar k.rapids.sq l.expressio n.Specified WindowFrame Sp ecification of the width of the group (or “frame”) of input rows around which a window function is evaluated true None
spark.rapi ds.sql.expr ession.Sqrt sqrt Square root true None
spark .rapids.sql .expression .StartsWith Starts with true None
spark. rapids.sql. expression. StringInstr instr Instr string operator true None
spark .rapids.sql .expression .StringLPad lpad Pad a string on the left true None
spark.r apids.sql.e xpression.S tringLocate `` position``, locate Substring search operator true None
spark .rapids.sql .expression .StringRPad rpad Pad a string on the right true None
spark.r apids.sql.e xpression.S tringRepeat repeat S tringRepeat operator that repeats the given strings with numbers of times given by repeatTimes true None
spark.ra pids.sql.ex pression.St ringReplace replace St ringReplace operator true None
spark. rapids.sql. expression. StringSplit split Splits str around occurrences that match regex true None
spark. rapids.sql. expression. StringToMap s tr_to_map Creates a map after splitting the input string into pairs of key-value strings true None
spark .rapids.sql .expression .StringTrim trim StringTrim operator true None
spark.rap ids.sql.exp ression.Str ingTrimLeft ltrim Str ingTrimLeft operator true None
spark.rapi ds.sql.expr ession.Stri ngTrimRight rtrim Stri ngTrimRight operator true None
spar k.rapids.sq l.expressio n.Substring substr, `` substring`` Substring operator true None
spark.rap ids.sql.exp ression.Sub stringIndex substr ing_index subs tring_index operator true None
spa rk.rapids.s ql.expressi on.Subtract - Subtraction true None
spark.rap ids.sql.exp ression.Tan tan Tangent true None
spark.rapi ds.sql.expr ession.Tanh tanh Hyperbolic tangent true None
sp ark.rapids. sql.express ion.TimeAdd Adds interval to timestamp true None
spar k.rapids.sq l.expressio n.ToDegrees degrees Converts radians to degrees true None
spar k.rapids.sq l.expressio n.ToRadians radians Converts degrees to radians true None
spark.rapi ds.sql.expr ession.ToUn ixTimestamp to_unix_ timestamp Returns the UNIX timestamp of the given time true None
spark.ra pids.sql.ex pression.Tr ansformKeys trans form_keys Transform keys in a map using a transform function true None
spark.rapi ds.sql.expr ession.Tran sformValues transfo rm_values Transform values in a map using a transform function true None
spark .rapids.sql .expression .UnaryMinus ` negative` Negate a numeric value true None
spark.ra pids.sql.ex pression.Un aryPositive ` positive` A numeric value with a + in front of it true None
spa rk.rapids.s ql.expressi on.Unbounde dFollowing$ Special boundary for a window frame, indicating all rows preceding the current row true None
spa rk.rapids.s ql.expressi on.Unbounde dPreceding$ Special boundary for a window frame, indicating all rows preceding the current row true None
spark.ra pids.sql.ex pression.Un ixTimestamp unix_ timestamp Returns the UNIX timestamp of current or specified time true None
spark.ra pids.sql.ex pression.Un scaledValue Convert a Decimal to an unscaled long value for some aggregation op timizations true None
spark.rapid s.sql.expre ssion.Upper upper, ucase String uppercase operator true This is not 100% compatible with the Spark version because the Unicode version used by cuDF and the JVM may differ, resulting in some corner-case characters not changing case correctly.
sp ark.rapids. sql.express ion.WeekDay weekday Returns the day of the week (0 = Monda y…6=Sunday) true None
spark.rapid s.sql.expre ssion.Windo wExpression Calculates a return value for every input row of a table based on a group (or “window”) of rows true None
spar k.rapids.sq l.expressio n.WindowSpe cDefinition Sp ecification of a window function, indicating the pa rtitioning- expression, the row ordering, and the width of the window true None
spark.rapi ds.sql.expr ession.Year year Returns the year from a date or timestamp true None
spa rk.rapids.s ql.expressi on.Aggregat eExpression Aggregate expression true None
spark .rapids.sql .expression .Approximat ePercentile percentil e_approx, approx_p ercentile Approximate percentile true This is not 100% compatible with the Spark version because the GPU imp lementation of approx _percentile is not bit-for-bit compatible with Apache Spark
sp ark.rapids. sql.express ion.Average avg, mean Average aggregate operator true None
spark. rapids.sql. expression. CollectList col lect_list Collect a list of non-unique elements, not supported in reduction true None
spark .rapids.sql .expression .CollectSet co llect_set Collect a set of unique elements, not supported in reduction true None
spark.rapid s.sql.expre ssion.Count count Count aggregate operator true None
spark.rapid s.sql.expre ssion.First fir st_value, first first aggregate operator true None
spark.rapi ds.sql.expr ession.Last last, l ast_value last aggregate operator true None
spark.rap ids.sql.exp ression.Max max Max aggregate operator true None
spark.rap ids.sql.exp ression.Min min Min aggregate operator true None
spark .rapids.sql .expression .PivotFirst PivotFirst operator true None
spar k.rapids.sq l.expressio n.StddevPop s tddev_pop Aggregation computing population standard deviation true None
spark .rapids.sql .expression .StddevSamp std dev_samp, std, stddev Aggregation computing sample standard deviation true None
spark.rap ids.sql.exp ression.Sum sum Sum aggregate operator true None
spark. rapids.sql. expression. VariancePop var_pop Aggregation computing population variance true None
spark.r apids.sql.e xpression.V arianceSamp `` var_samp``, ` variance` Aggregation computing sample variance true None
spa rk.rapids.s ql.expressi on.Normaliz eNaNAndZero Normalize NaN and zero true None
spark.rap ids.sql.exp ression.Sca larSubquery Subquery that will return only one row and one column true None
spark.rap ids.sql.exp ression.Hiv eGenericUDF Hive Generic UDF, the UDF can choose to implement a RAPIDS accelerated interface to get better performance true None
spark.ra pids.sql.ex pression.Hi veSimpleUDF Hive UDF, the UDF can choose to implement a RAPIDS accelerated interface to get better performance true None

Execution

Name

Description

Default Value

Notes

spark. rapids.sql.exe c.CoalesceExec The backend for the dataframe coalesce method true None
spark.rapi ds.sql.exec.Co llectLimitExec Reduce to single partition and apply limit false This is disabled by default because Collect Limit replacement can be slower on the GPU, if huge number of rows in a batch it could help by limiting the number of rows transferred from GPU to CPU
spar k.rapids.sql.e xec.ExpandExec The backend for the expand operator true None
spark.rapids .sql.exec.File SourceScanExec Reading data from files, often from Hive tables true None
spar k.rapids.sql.e xec.FilterExec The backend for most filter statements true None
spark. rapids.sql.exe c.GenerateExec The backend for operations that generate more output rows than input rows like explode true None
spark.rap ids.sql.exec.G lobalLimitExec Limiting of results across partitions true None
spark.ra pids.sql.exec. LocalLimitExec Per-partition limiting of results true None
spark .rapids.sql.ex ec.ProjectExec The backend for most select, withColumn and dropColumn statements true None
spa rk.rapids.sql. exec.RangeExec The backend for range operator true None
spar k.rapids.sql.e xec.SampleExec The backend for the sample operator true None
sp ark.rapids.sql .exec.SortExec The backend for the sort operator true None
s park.rapids.sq l.exec.Subquer yBroadcastExec Plan to collect and transform the broadcast key values true None
spark .rapids.sql.ex ec.TakeOrdered AndProjectExec Take the first limit elements as defined by the sortOrder, and do projection if needed true None
spa rk.rapids.sql. exec.UnionExec The backend for the union operator true None
spa rk.rapids.sql. exec.CustomShu ffleReaderExec A wrapper of shuffle query stage true None
spark.rapid s.sql.exec.Has hAggregateExec The backend for hash based aggregations true None
spa rk.rapids.sql. exec.ObjectHas hAggregateExec The backend for hash based aggregations supporting TypedImper ativeAggregate functions true None
spark.rapid s.sql.exec.Sor tAggregateExec The backend for sort based aggregations true None
s park.rapids.sq l.exec.InMemor yTableScanExec Implementation of InMemor yTableScanExec to use GPU accelerated caching true None
sp ark.rapids.sql .exec.DataWrit ingCommandExec Writing data true None
spark.rapids. sql.exec.Execu tedCommandExec Eagerly executed commands true None
spark.r apids.sql.exec .BatchScanExec The backend for most file input true None
s park.rapids.sq l.exec.Broadca stExchangeExec The backend for broadcast exchange of data true None
spark.rapids. sql.exec.Shuff leExchangeExec The backend for most data being exchanged between processes true None
s park.rapids.sq l.exec.Broadca stHashJoinExec Implementation of join using broadcast data true None
spark.r apids.sql.exec .BroadcastNest edLoopJoinExec Implementation of join using brute force. Full outer joins and joins where the broadcast side matches the join side (e.g.: LeftOuter with left broadcast) are not supported true None
spark.rapids.s ql.exec.Cartes ianProductExec Implementation of join using brute force true None
spark.rapids.s ql.exec.Shuffl edHashJoinExec Implementation of join using hashed shuffled data true None
spark.rapid s.sql.exec.Sor tMergeJoinExec Sort merge join, replacing with shuffled hash join true None
s park.rapids.sq l.exec.Aggrega teInPandasExec The backend for an Aggregation Pandas UDF, this accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled. true None
spark.rapids. sql.exec.Arrow EvalPythonExec The backend of the Scalar Pandas UDFs. Accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled true None
spark.r apids.sql.exec .FlatMapCoGrou psInPandasExec The backend for CoGrouped Aggregation Pandas UDF. Accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled. false This is disabled by default because Performance is not ideal with many small groups
spark .rapids.sql.ex ec.FlatMapGrou psInPandasExec The backend for Flat Map Groups Pandas UDF, Accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled. true None
spark.rap ids.sql.exec.M apInPandasExec The backend for Map Pandas Iterator UDF. Accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled. true None
spark.rapids .sql.exec.Wind owInPandasExec The backend for Window Aggregation Pandas UDF, Accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled. For now it only supports row based window frame. false This is disabled by default because it only supports row based frame for now
spar k.rapids.sql.e xec.WindowExec W indow-operator backend true None
spark.rapid s.sql.exec.Hiv eTableScanExec Scan Exec to read Hive delimited text tables true None

Commands

Name

Description

Default Value

Notes

spark.rapids .sql.command.SaveIn toDataSourceCommand Write to a data source true None

Scans

Name

Description

Default Value

Notes

spark.rapids.sql.input.CSVScan CSV parsing true None
spark.rapids.sql.input.JsonScan Json parsing true None
spark.rapids.sql.input.OrcScan ORC parsing true None
spark.rapids.sql.input.ParquetScan Parquet parsing true None
spark.rapids.sql.input.AvroScan Avro parsing true None

Partitioning

Name

Description

Default Value

Notes

spark.ra pids.sql.partitioni ng.HashPartitioning Hash based partitioning true None
spark.rap ids.sql.partitionin g.RangePartitioning Range partitioning true None
spark.rapids.s ql.partitioning.Rou ndRobinPartitioning Round robin partitioning true None
spark.ra pids.sql.partitioni ng.SinglePartition$ Single partitioning true None
Previous Additional Resources
Next RAPIDS Accelerator for Apache Spark Compatibility with Apache Spark
© Copyright 2024, NVIDIA. Last updated on Apr 2, 2024.