DOCA Telemetry Service Guide

This guide provides instructions on how to use the DOCA Telemetry Service (DTS) container on top of hosts or NVIDIA® BlueField® DPU.

DOCA Telemetry Service (DTS) collects data from built-in providers and from external telemetry applications.

DTS supports the following providers:

• Data providers:

  • Sysfs

  • Ethtool

  • Advanced monitoring and bringup errors (amber)

  • Port programmable congestion control (ppcc_eth)

  • Diagnostic data

  • Ifconfig

  • RDMA notifications

  • Data center GPU manager (DCGM)

  • NVIDIA System Management Interface (nvidia-smi)

  • BlueField performance (bfperf)

  • TC (traffic control)

• Aggregation providers:

  • Fluent aggregator

  • Prometheus aggregator

Data Storage

DTS stores collected data in binary files located at /opt/mellanox/doca/services/telemetry/data directory. By default, data writing is disabled on BlueField due to storage limitations.

Data Export Options

DTS supports the following export mechanisms:

• Prometheus Endpoint (pull)

• Fluent Bit (push)

• OpenTelemetry (push)

• Prometheus Remote Write (push)

NetFlow Packet Export

DTS can export NetFlow packets when data is collected using the DOCA Telemetry Exporter NetFlow API client application.

To enable the NetFlow exporter, configure the dts_config.ini file with the NetFlow collector IP/address and port.

Available Images

Built-in DOCA Service Image

DOCA Telemetry Service is enabled by default on the DPU and is shipped as part of the BlueField image. That is, every BlueField image contains a fixed service version so as to provide out-of-the-box support for programs based on the 2024-10-09_07-10-18_DOCA Telemetry library.

DOCA Service on NGC

In addition to the built-in image shipped with the BlueField boot image, DTS is also available on NGC, NVIDIA's container catalog. This is useful in case a new version of the service has been released and the user wants to upgrade from the built-in image. For service-specific configuration steps and deployment instructions, refer to the service's container page .

DPU Deployment

As mentioned above, DTS starts automatically on BlueField boot. This is done according to the .yaml file located at /etc/kubelet.d/doca_telemetry_standalone.yaml. Removing the .yaml file from this path stops the automatic DTS boot.

DTS files can be found under the directory /opt/mellanox/doca/services/telemetry/.

• Container folder mounts:

  • config

  • data

  • ipc_sockets

• Backup files:

  • doca_telemetry_service_${version}_arm64.tar.gz – DTS image

  • doca_telemetry_standalone.yaml – copy of the default boot .yaml file

Host Deployment

DTS supports x86_64 hosts. The providers and exporters all run from a single docker container.

1. Initialize and configure host DTS with the desired DTS version:

   Copy

   Copied!

               
   
               
   export DTS_IMAGE=nvcr.io/nvidia/doca/doca_telemetry:<desired-DTS-version>
   docker run -v "/opt/mellanox/doca/services/telemetry/config:/config" --rm --name doca-telemetry-init -it $DTS_IMAGE /bin/bash -c "DTS_CONFIG_DIR=host /usr/bin/telemetry-init.sh"
           

   Note

   Per NGC policy, the "latest" tag does not exist. This means that when deploying DTS, the user must pick the desired tag from NGC and ensure that the DTS_IMAGE variable points to the full image. Example from version 1.16.5-doca2.6.0-host:

   Copy

   Copied!

               
   
               
   export DTS_IMAGE=nvcr.io/nvidia/doca/doca_telemetry:1.16.5-doca2.6.0-host
           

2. Run with:

   Copy

   Copied!

               
   
               
   docker run -d --net=host --uts=host --ipc=host                                        \
                 --privileged                                                            \
                 --ulimit stack=67108864 --ulimit memlock=-1                             \
                 --device=/dev/mst/                                                      \
                 --device=/dev/infiniband/                                               \
                 --gpus all                                                              \
                 -v "/opt/mellanox/doca/services/telemetry/config:/config"               \
                 -v "/opt/mellanox/doca/services/telemetry/ipc_sockets:/tmp/ipc_sockets" \
                 -v "/opt/mellanox/doca/services/telemetry/data:/data"                   \
                 -v "/usr/lib/mft:/usr/lib/mft"                                          \
                 -v "/sys/kernel/debug:/sys/kernel/debug"                                \
                 --rm --name doca-telemetry -it $DTS_IMAGE /usr/bin/telemetry-run.sh
           

   Note

   The following mounts are required by specific services only:

   • hcaperf provider:

     • --device=/dev/mst/

     • -v "/usr/lib/mft:/usr/lib/mft"

     • -v "/sys/kernel/debug:/sys/kernel/debug"

   • UCX/RDMA export modes:

     • --device=/dev/infiniband/

   • GPU providers (nvidia-smi and dcgm):

     • --gpu all

Deployment with Grafana Monitoring

Refer to section "Deploying with Grafana Monitoring".

The configuration of DTS is placed under /opt/mellanox/doca/services/telemetry/config by DTS during initialization. The user can interact with the dts_config.ini file and fluent_bit_configs folder. dts_config.ini contains the main configuration for the service and must be used to enable/disable providers, exporters, data writing. More details are provided in the corresponding sections. For every update in this file, DST must be restarted. Interaction with fluent_bit_configs folder is described in section Fluent Bit.

Init Scripts

The InitContainers section of the .yaml file has 2 scripts for config initialization:

• /usr/bin/telemetry-init.sh – generates the default configuration files if, and only if, the /opt/mellanox/doca/services/telemetry/config folder is empty.

• /usr/bin/enable-fluent-forward.sh – configures the destination host and port for Fluent Bit forwarding. The script requires that both the host and port are present, and only in this case it would start. The script overwrites the /opt/mellanox/doca/services/telemetry/config/fluent_bit_configs folder and configures the .exp file.

Enabling Fluent Bit Forwarding

To enable Fluent Bit forward, add the destination host and port to the command line found in the initContainers section of the .yaml file:

            

            
command: ["/bin/bash", "-c", "/usr/bin/telemetry-init.sh && /usr/bin/enable-fluent-forward.sh -i=127.0.0.1 -p=24224"]
        

Generating Configuration

The configuration folder /opt/mellanox/doca/services/telemetry/config starts empty by default. Once the service starts, the initial scripts run as a part of the initial container and create configuration as described in section Enabling Fluent Bit Forwarding.

Resetting Configuration

Resetting the configuration can be done by deleting the content found in the configuration folder and restarting the service to generate the default configuration.

Sampling Interval

The sampling interval is shared among all telemetry providers (with few exceptions) and is controlled by the update parameter:

            

            
# Sampling interval for providers in milliseconds
update=1000
        

The sampling frequency may not be reflected in the export side, as the collected data is exported only when an internal buffer (also called data page) is filled with data. This may result in the DTS exporter update frequency to be lower than the providers sampling frequency.

That is, if a data page contains N counters and a provider fills up 10 counters on each sample, it would take N/10 samples to fill up the data page and export the data. This logic can be overridden with the sync-time-limit parameter which defines the maximum delay time between collected and exported data:

            

            
# Timeout for forced page rotation in seconds
sync-time-limit=10000
        

Enabling Providers

Providers are enabled from the dts_config.ini configuration file. Uncomment the enable-provider=$provider-name line to allow data collection for this provider. For example, uncommenting the following line enables the ethtool provider:

            

            
#enable-provider=ethtool
        

Remote Collection

Certain providers or components are unable to execute properly within the container due to various container limitations. Therefore, they would have to perform remote collection or execution.

The following steps enable remote collection:

1. Activate DOCA privileged executer (DPE), as DPE is how remote collection is achieved:

   Copy

   Copied!

               
   
               
   systemctl start dpe
           

2. Add grpc before provider-name (i.e., enable-provider=grpc.$provider-name). For example, the following line configures remote collection of the hcaperf provider:

   Copy

   Copied!

               
   
               
   enable-provider=grpc.hcaperf
           

3. If there are any configuration lines that are provider-specific, then add the grpc prefix as well. Building upon the previous example:

   Copy

   Copied!

               
   
               
   grpc.hcaperf.mlx5_0=sample
   grpc.hcaperf.mlx5_1=sample 
           

Enabling Data Write

Uncomment the following line in dts_config.ini:

            

            
#output=/data
        

Enabling IPC with Non-container Program

For information on enabling IPC between DTS and an application that runs outside of a container, refer to section "Using IPC with Non-container Application" in DOCA Telemetry Exporter.

Level Labels

For DTS exporters that support metadata labels, such as Prometheus and OpenTelemetry, labels can be specified for two levels of telemetry: global and per-device. These labels allow for precise identification and categorization of telemetry data.

Configuring Level Labels

The level labels are defined in the level_labels.ini file, located at /opt/mellanox/doca/telemetry/config/level_labels.ini.

The following is an example level_labels.ini file, defining global and per-device labels:

            

            
# lines starting with hash are ignored
# 2 level labels are supported - global and device
# global labels will be added to all metrics
# device labels will be added to metrics matching the device description
 
[global_labels]
# global labels to be added to all metrics
# format:
# label_name=label_value
hostname=myhost
ip=127.127.127.127
 
[device_labels]
# device labels will be added to metrics related to the device
# format:
# device_index_label=device_label_name1|device_label_value1|device_label_name2|device_label_value2
mlx5_0=rail|0|network|compute
mlx5_1=rail|1|network|compute
mlx5_2=rail|2|network|storage
 
[device_mapping]
# optional: mapping of device names to network interfaces and mst devices
# this is done internally by DTS, but can be overridden here
# add clear_auto_detected_mapping=true to clear the auto-detected mapping
# format:
# device_name=mst_device_name|net_interface_names
# in case of multiple net interfaces, separate them with comma
 
[data_types_mapping]
# internal data types, don't change this section if you don't know what you are doing
# format:
# data_type|index|device_type_name
ethtool_event=device_name|netif
ppcc_eth=device_name|mst
ifconfig_event=device_name|netif
        

Enabling Level Labels

1. Ensure the global_labels and device_labels sections are correctly configured in the level_labels.ini file.

2. Confirm the DTS configuration file (/opt/mellanox/doca/telemetry/config/dts_config.ini) includes the following line:

   Copy

   Copied!

               
   
               
   level-labels-file=/config/level_labels.ini
           

Monitoring and Reloading

When enabled, DTS monitors the level_labels.ini file. If the file is modified:

• The telemetry collection process in DTS will automatically exit

• DTS will restart after 60 seconds, allowing time for the updated configuration to take effect

This ensures that changes to the level labels file are applied without requiring manual restarts.

Providers

DTS supports on-board data collection from sysf, ethtool, and tc providers. Fluent and Prometheus aggregator providers can collect the data from other applications.

Other providers are available based on different conditions (e.g., specific container mounts or host only such as amber, ppcc_eth, etc). Such providers are described with their dependencies in their corresponding sections.

Sysfs Counters List

The sysfs provider has several components: ib_port, hw_port, mr_cache, eth, hwmon and bf_ptm . By default, all the components (except bf_ptm) are enabled when the provider is enabled:

            

            
#disable-provider=sysfs
        

The components can be disabled separately. For instance, to disable eth:

            

            
enable-provider=sysfs
disable-provider=sysfs.eth
        

• ib_port counters:

  Copy

  Copied!

              
  
              
  {hca_name}:{port_num}:ib_port_state
  {hca_name}:{port_num}:VL15_dropped
  {hca_name}:{port_num}:excessive_buffer_overrun_errors
  {hca_name}:{port_num}:link_downed
  {hca_name}:{port_num}:link_error_recovery
  {hca_name}:{port_num}:local_link_integrity_errors
  {hca_name}:{port_num}:multicast_rcv_packets
  {hca_name}:{port_num}:multicast_xmit_packets
  {hca_name}:{port_num}:port_rcv_constraint_errors
  {hca_name}:{port_num}:port_rcv_data
  {hca_name}:{port_num}:port_rcv_errors
  {hca_name}:{port_num}:port_rcv_packets
  {hca_name}:{port_num}:port_rcv_remote_physical_errors
  {hca_name}:{port_num}:port_rcv_switch_relay_errors
  {hca_name}:{port_num}:port_xmit_constraint_errors
  {hca_name}:{port_num}:port_xmit_data
  {hca_name}:{port_num}:port_xmit_discards
  {hca_name}:{port_num}:port_xmit_packets
  {hca_name}:{port_num}:port_xmit_wait
  {hca_name}:{port_num}:symbol_error
  {hca_name}:{port_num}:unicast_rcv_packets
  {hca_name}:{port_num}:unicast_xmit_packets
          

• ib_hw counters:

  Copy

  Copied!

              
  
              
  {hca_name}:{port_num}:hw_state
  {hca_name}:{port_num}:hw_duplicate_request
  {hca_name}:{port_num}:hw_implied_nak_seq_err
  {hca_name}:{port_num}:hw_lifespan
  {hca_name}:{port_num}:hw_local_ack_timeout_err
  {hca_name}:{port_num}:hw_out_of_buffer
  {hca_name}:{port_num}:hw_out_of_sequence
  {hca_name}:{port_num}:hw_packet_seq_err   
  {hca_name}:{port_num}:hw_req_cqe_error
  {hca_name}:{port_num}:hw_req_cqe_flush_error
  {hca_name}:{port_num}:hw_req_remote_access_errors   
  {hca_name}:{port_num}:hw_req_remote_invalid_request   
  {hca_name}:{port_num}:hw_resp_cqe_error
  {hca_name}:{port_num}:hw_resp_cqe_flush_error
  {hca_name}:{port_num}:hw_resp_local_length_error
  {hca_name}:{port_num}:hw_resp_remote_access_errors
  {hca_name}:{port_num}:hw_rnr_nak_retry_err   
  {hca_name}:{port_num}:hw_rx_atomic_requests   
  {hca_name}:{port_num}:hw_rx_dct_connect
  {hca_name}:{port_num}:hw_rx_icrc_encapsulated
  {hca_name}:{port_num}:hw_rx_read_requests   
  {hca_name}:{port_num}:hw_rx_write_requests
          

• ib_mr_cache counters:

  Copy

  Copied!

              
  
              
  {hca_name}:mr_cache:size_{n}:cur
  {hca_name}:mr_cache:size_{n}:limit
  {hca_name}:mr_cache:size_{n}:miss
  {hca_name}:mr_cache:size_{n}:size
          

  Note

  Where n ranges from 0 to 24.

• eth counters:

  Copy

  Copied!

              
  
              
  {hca_name}:{device_name}:eth_collisions
  {hca_name}:{device_name}:eth_multicast
  {hca_name}:{device_name}:eth_rx_bytes
  {hca_name}:{device_name}:eth_rx_compressed
  {hca_name}:{device_name}:eth_rx_crc_errors
  {hca_name}:{device_name}:eth_rx_dropped
  {hca_name}:{device_name}:eth_rx_errors
  {hca_name}:{device_name}:eth_rx_fifo_errors
  {hca_name}:{device_name}:eth_rx_frame_errors
  {hca_name}:{device_name}:eth_rx_length_errors
  {hca_name}:{device_name}:eth_rx_missed_errors
  {hca_name}:{device_name}:eth_rx_nohandler
  {hca_name}:{device_name}:eth_rx_over_errors
  {hca_name}:{device_name}:eth_rx_packets
  {hca_name}:{device_name}:eth_tx_aborted_errors
  {hca_name}:{device_name}:eth_tx_bytes
  {hca_name}:{device_name}:eth_tx_carrier_errors
  {hca_name}:{device_name}:eth_tx_compressed
  {hca_name}:{device_name}:eth_tx_dropped
  {hca_name}:{device_name}:eth_tx_errors
  {hca_name}:{device_name}:eth_tx_fifo_errors
  {hca_name}:{device_name}:eth_tx_heartbeat_errors
  {hca_name}:{device_name}:eth_tx_packets
  {hca_name}:{device_name}:eth_tx_window_errors
          

• devices counters

  Copy

  Copied!

              
  
              
  {hca_name}:current_link_width
  {hca_name}:current_link_speed
  {hca_name}:max_link_speed
  {hca_name}:max_link_width
          

• BlueField-2 hwmon counters:

  Collapse Source

  Copy

  Copied!

              
  
              
  {hwmon_name}:{l3cache}:CYCLES
  {hwmon_name}:{l3cache}:HITS_BANK0
  {hwmon_name}:{l3cache}:HITS_BANK1
  {hwmon_name}:{l3cache}:MISSES_BANK0
  {hwmon_name}:{l3cache}:MISSES_BANK1
  {hwmon_name}:{pcie}:IN_C_BYTE_CNT
  {hwmon_name}:{pcie}:IN_C_PKT_CNT
  {hwmon_name}:{pcie}:IN_NP_BYTE_CNT
  {hwmon_name}:{pcie}:IN_NP_PKT_CNT
  {hwmon_name}:{pcie}:IN_P_BYTE_CNT
  {hwmon_name}:{pcie}:IN_P_PKT_CNT
  {hwmon_name}:{pcie}:OUT_C_BYTE_CNT
  {hwmon_name}:{pcie}:OUT_C_PKT_CNT
  {hwmon_name}:{pcie}:OUT_NP_BYTE_CNT
  {hwmon_name}:{pcie}:OUT_NP_PKT_CNT
  {hwmon_name}:{pcie}:OUT_P_PKT_CNT
  {hwmon_name}:{tile}:MEMORY_READS
  {hwmon_name}:{tile}:MEMORY_WRITES
  {hwmon_name}:{tile}:MSS_NO_CREDIT
  {hwmon_name}:{tile}:VICTIM_WRITE
  {hwmon_name}:{tilenet}:CDN_DIAG_C_OUT_OF_CRED
  {hwmon_name}:{tilenet}:CDN_REQ
  {hwmon_name}:{tilenet}:DDN_REQ
  {hwmon_name}:{tilenet}:NDN_REQ
  {hwmon_name}:{trio}:TDMA_DATA_BEAT
  {hwmon_name}:{trio}:TDMA_PBUF_MAC_AF
  {hwmon_name}:{trio}:TDMA_RT_AF
  {hwmon_name}:{trio}:TPIO_DATA_BEAT
  {hwmon_name}:{triogen}:TX_DAT_AF
  {hwmon_name}:{triogen}:TX_DAT_AF
          

• BlueField-3 hwmon counters:

  Copy

  Copied!

              
  
              
  {hwmon_name}:{llt}:GDC_BANK0_RD_REQ
  {hwmon_name}:{llt}:GDC_BANK1_RD_REQ
  {hwmon_name}:{llt}:GDC_BANK0_WR_REQ
  {hwmon_name}:{llt}:GDC_BANK1_WR_REQ
  {hwmon_name}:{llt_miss}:GDC_MISS_MACHINE_RD_REQ
  {hwmon_name}:{llt_miss}:GDC_MISS_MACHINE_WR_REQ
  {hwmon_name}:{mss}:SKYLIB_DDN_TX_FLITS
  {hwmon_name}:{mss}:SKYLIB_DDN_RX_FLITS
          

• BlueField-3 bf_ptm counters:

  Copy

  Copied!

              
  
              
  bf:ptm:active_power_profile
  bf:ptm:atx_power_available
  bf:ptm:core_temp
  bf:ptm:ddr_temp
  bf:ptm:error_state
  bf:ptm:power_envelope
  bf:ptm:power_throttling_event_count
  bf:ptm:power_throttling_state
  bf:ptm:thermal_throttling_event_count
  bf:ptm:thermal_throttling_state
  bf:ptm:throttling_state
  bf:ptm:total_power
  bf:ptm:vr0_power
  bf:ptm:vr1_power
          

• rate counters – calculated counters showing the rate of raw counters of other components. Such counters are identified by a _rate suffix which correspond to the original raw counter name they track.

Port Counters

The following parameters are located in /sys/class/infiniband/mlx5_0/ports/1/counters.

Hardware Counters

The hardware counters, found under /sys/class/infiniband/mlx5_0/ports/1/hw_counters/, are counted per function and exposed on the function. Some counters are not counted per function. These counters are commented with a relevant comment.

Debug Status Counters

The following parameters are located in /sys/class/net/<interface>/debug.

            

            
# cat /sys/class/net/eth2/debug/lro_timeout
Actual timeout: 32
Supported timeout: 8 16 32 1024
        

            

            
$ cat /sys/class/net/ethXX/debug/link_down_reason
monitor_opcode: 0x0
status_message: The port is Active.
        

Power Thermal Counters

The bf_ptm component collects BlueField-3 power thermal counters using remote collection. It is disabled by default and can be enabled as follows:

1. Load kernel module mlxbf-ptm:

   Copy

   Copied!

               
   
               
   modprobe -v mlxbf-ptm
           

2. Enable component using remote collection:

   Copy

   Copied!

               
   
               
   enable-provider=grpc.sysfs.bf_ptm
           

   Note

   DPE server should be active before changing the dts_config.ini file. See section "Remote Collection" for details.

Ethtool Counters

Ethtool counters is the generated list of counters which corresponds to Ethtool utility. Counters are generated on a per-device basis.

There are several counter groups, depending on where the counter is counted:

• Ring – software ring counters

• Software port – an aggregation of software ring counters

• vPort counters – traffic counters and drops due to steering or no buffers. May indicate BlueField issues. These counters include Ethernet traffic counters (including raw Ethernet) and RDMA/RoCE traffic counters.

• Physical port counters – the physical port connecting BlueField to the network. May indicate device issues or link or network issues. This measuring point holds information on standardized counters like IEEE 802.3, RFC2863, RFC 2819, RFC 3635 and additional counters like flow control, FEC, and more. Physical port counters are not exposed to virtual machines.

• Priority port counters – a set of the physical port counters, per priory per port

Each group of counters may have different counter types:

• Traffic informative counters – counters which counts traffic. These counters can be used for load estimation of for general debug.

• Traffic acceleration counters – counters which counts traffic accelerated by NVIDIA drivers or by hardware. The counters are an additional layer to the informative counter set and the same traffic is counted in both informative and acceleration counters. Acceleration counters are marked with [A].

• Error counters – increment of these counters might indicate a problem

The following acceleration mechanisms have dedicated counters:

• TCP segmentation offload (TSO) – increasing outbound throughput and reducing CPU utilization by allowing the kernel to buffer multiple packets in a single large buffer. The BlueField splits the buffer into packet and transmits it.

• Large receive offload (LRO) – increasing inbound throughput and reducing CPU utilization by aggregation of multiple incoming packets of a single stream to a single buffer

• CHECKSUM – calculation of TCP checksum (by the BlueField). The following checksum offloads are available ( refer to skbuff.h for detailed explanation)

  • CHECKSUM_UNNECESSARY

  • CHECKSUM_NONE – no checksum acceleration was used

  • CHECKSUM_COMPLETE – device provided checksum on the entire packet

  • CHECKSUM_PARTIAL – device provided checksum

• CQE compress – compression of completion queue events (CQE) used for sparing bandwidth on PCIe and hence achieve better performance.

Ring/Software Port Counters

The following counters are available per ring or software port.

These counters provide information on the amount of traffic accelerated by the BlueField. The counters tally the accelerated traffic in addition to the standard counters which tally that (i.e. accelerated traffic is counted twice).

The counter names in the table below refers to both ring and port counters. the notation for ring counters includes the [i] index without the braces. the notation for port counters does not include the [i]. a counter name rx[i]_packets will be printed as rx0_packets for ring 0 and rx_packets for the software port

1. The corresponding ring and global counters do not share the same name (i.e., do not follow the common naming scheme).                          

vPort Counters

Counters on the eswitch port that is connected to the vNIC.

Physical Port Counters

The physical port counters are the counters on the external port connecting adapter to the network. This measuring point holds information on standardized counters like IEEE 802.3, RFC2863, RFC 2819, RFC 3635 and additional counters like flow control, FEC and more.

Priority Port Counters

The following counters are physical port counters that being counted per L2 priority (0-7).

Device Counters

Full List of Counters

            

            
# ethtool -S eth5
 
NIC statistics:
rx_packets: 10
rx_bytes: 3420
tx_packets: 18
tx_bytes: 1296
tx_tso_packets: 0
tx_tso_bytes: 0
tx_tso_inner_packets: 0
tx_tso_inner_bytes: 0
tx_added_vlan_packets: 0
tx_nop: 0
rx_lro_packets: 0
rx_lro_bytes: 0
rx_ecn_mark: 0
rx_removed_vlan_packets: 0
rx_csum_unnecessary: 0
rx_csum_none: 0
rx_csum_complete: 10
rx_csum_unnecessary_inner: 0
rx_xdp_drop: 0
rx_xdp_redirect: 0
rx_xdp_tx_xmit: 0
rx_xdp_tx_full: 0
rx_xdp_tx_err: 0
rx_xdp_tx_cqe: 0
tx_csum_none: 18
tx_csum_partial: 0
tx_csum_partial_inner: 0
tx_queue_stopped: 0
tx_queue_dropped: 0
tx_xmit_more: 0
tx_recover: 0
tx_cqes: 18
tx_queue_wake: 0
tx_udp_seg_rem: 0
tx_cqe_err: 0
tx_xdp_xmit: 0
tx_xdp_full: 0
tx_xdp_err: 0
tx_xdp_cqes: 0
rx_wqe_err: 0
rx_mpwqe_filler_cqes: 0
rx_mpwqe_filler_strides: 0
rx_buff_alloc_err: 0
rx_cqe_compress_blks: 0
rx_cqe_compress_pkts: 0
rx_page_reuse: 0
rx_cache_reuse: 0
rx_cache_full: 0
rx_cache_empty: 2688
rx_cache_busy: 0
rx_cache_waive: 0
rx_congst_umr: 0
rx_arfs_err: 0
ch_events: 75
ch_poll: 75
ch_arm: 75
ch_aff_change: 0
ch_eq_rearm: 0
rx_out_of_buffer: 0
rx_if_down_packets: 15
rx_steer_missed_packets: 0
rx_vport_unicast_packets: 0
rx_vport_unicast_bytes: 0
tx_vport_unicast_packets: 0
tx_vport_unicast_bytes: 0
rx_vport_multicast_packets: 2
rx_vport_multicast_bytes: 172
tx_vport_multicast_packets: 12
tx_vport_multicast_bytes: 936
rx_vport_broadcast_packets: 37
rx_vport_broadcast_bytes: 9270
tx_vport_broadcast_packets: 6
tx_vport_broadcast_bytes: 360
rx_vport_rdma_unicast_packets: 0
rx_vport_rdma_unicast_bytes: 0
tx_vport_rdma_unicast_packets: 0
tx_vport_rdma_unicast_bytes: 0
rx_vport_rdma_multicast_packets: 0
rx_vport_rdma_multicast_bytes: 0
tx_vport_rdma_multicast_packets: 0
tx_vport_rdma_multicast_bytes: 0
tx_packets_phy: 0
rx_packets_phy: 0
rx_crc_errors_phy: 0
tx_bytes_phy: 0
rx_bytes_phy: 0
tx_multicast_phy: 0
tx_broadcast_phy: 0
rx_multicast_phy: 0
rx_broadcast_phy: 0
rx_in_range_len_errors_phy: 0
rx_out_of_range_len_phy: 0
rx_oversize_pkts_phy: 0
rx_symbol_err_phy: 0
tx_mac_control_phy: 0
rx_mac_control_phy: 0
rx_unsupported_op_phy: 0
rx_pause_ctrl_phy: 0
tx_pause_ctrl_phy: 0
rx_discards_phy: 0
tx_discards_phy: 0
tx_errors_phy: 0
rx_undersize_pkts_phy: 0
rx_fragments_phy: 0
rx_jabbers_phy: 0
rx_64_bytes_phy: 0
rx_65_to_127_bytes_phy: 0
rx_128_to_255_bytes_phy: 0
rx_256_to_511_bytes_phy: 0
rx_512_to_1023_bytes_phy: 0
rx_1024_to_1518_bytes_phy: 0
rx_1519_to_2047_bytes_phy: 0
rx_2048_to_4095_bytes_phy: 0
rx_4096_to_8191_bytes_phy: 0
rx_8192_to_10239_bytes_phy: 0
link_down_events_phy: 0
rx_prio0_bytes: 0
rx_prio0_packets: 0
tx_prio0_bytes: 0
tx_prio0_packets: 0
rx_prio1_bytes: 0
rx_prio1_packets: 0
tx_prio1_bytes: 0
tx_prio1_packets: 0
rx_prio2_bytes: 0
rx_prio2_packets: 0
tx_prio2_bytes: 0
tx_prio2_packets: 0
rx_prio3_bytes: 0
rx_prio3_packets: 0
tx_prio3_bytes: 0
tx_prio3_packets: 0
rx_prio4_bytes: 0
rx_prio4_packets: 0
tx_prio4_bytes: 0
tx_prio4_packets: 0
rx_prio5_bytes: 0
rx_prio5_packets: 0
tx_prio5_bytes: 0
tx_prio5_packets: 0
rx_prio6_bytes: 0
rx_prio6_packets: 0
tx_prio6_bytes: 0
tx_prio6_packets: 0
rx_prio7_bytes: 0
rx_prio7_packets: 0
tx_prio7_bytes: 0
tx_prio7_packets: 0
module_unplug: 0
module_bus_stuck: 0
module_high_temp: 0
module_bad_shorted: 0
ch0_events: 9
ch0_poll: 9
ch0_arm: 9
ch0_aff_change: 0
ch0_eq_rearm: 0
ch1_events: 23
ch1_poll: 23
ch1_arm: 23
ch1_aff_change: 0
ch1_eq_rearm: 0
ch2_events: 8
ch2_poll: 8
ch2_arm: 8
ch2_aff_change: 0
ch2_eq_rearm: 0
ch3_events: 19
ch3_poll: 19
ch3_arm: 19
ch3_aff_change: 0
ch3_eq_rearm: 0
ch4_events: 8
ch4_poll: 8
ch4_arm: 8
ch4_aff_change: 0
ch4_eq_rearm: 0
ch5_events: 8
ch5_poll: 8
ch5_arm: 8
ch5_aff_change: 0
ch5_eq_rearm: 0
rx0_packets: 0
rx0_bytes: 0
rx0_csum_complete: 0
rx0_csum_unnecessary: 0
rx0_csum_unnecessary_inner: 0
rx0_csum_none: 0
rx0_xdp_drop: 0
rx0_xdp_redirect: 0
rx0_lro_packets: 0
rx0_lro_bytes: 0
rx0_ecn_mark: 0
rx0_removed_vlan_packets: 0
rx0_wqe_err: 0
rx0_mpwqe_filler_cqes: 0
rx0_mpwqe_filler_strides: 0
rx0_buff_alloc_err: 0
rx0_cqe_compress_blks: 0
rx0_cqe_compress_pkts: 0
rx0_page_reuse: 0
rx0_cache_reuse: 0
rx0_cache_full: 0
rx0_cache_empty: 448
rx0_cache_busy: 0
rx0_cache_waive: 0
rx0_congst_umr: 0
rx0_arfs_err: 0
rx0_xdp_tx_xmit: 0
rx0_xdp_tx_full: 0
rx0_xdp_tx_err: 0
rx0_xdp_tx_cqes: 0
rx1_packets: 10
rx1_bytes: 3420
rx1_csum_complete: 10
rx1_csum_unnecessary: 0
rx1_csum_unnecessary_inner: 0
rx1_csum_none: 0
rx1_xdp_drop: 0
rx1_xdp_redirect: 0
rx1_lro_packets: 0
rx1_lro_bytes: 0
rx1_ecn_mark: 0
rx1_removed_vlan_packets: 0
rx1_wqe_err: 0
rx1_mpwqe_filler_cqes: 0
rx1_mpwqe_filler_strides: 0
rx1_buff_alloc_err: 0
rx1_cqe_compress_blks: 0
rx1_cqe_compress_pkts: 0
rx1_page_reuse: 0
rx1_cache_reuse: 0
rx1_cache_full: 0
rx1_cache_empty: 448
rx1_cache_busy: 0
rx1_cache_waive: 0
rx1_congst_umr: 0
rx1_arfs_err: 0
rx1_xdp_tx_xmit: 0
rx1_xdp_tx_full: 0
rx1_xdp_tx_err: 0
rx1_xdp_tx_cqes: 0
rx2_packets: 0
rx2_bytes: 0
rx2_csum_complete: 0
rx2_csum_unnecessary: 0
rx2_csum_unnecessary_inner: 0
rx2_csum_none: 0
rx2_xdp_drop: 0
rx2_xdp_redirect: 0
rx2_lro_packets: 0
rx2_lro_bytes: 0
rx2_ecn_mark: 0
rx2_removed_vlan_packets: 0
rx2_wqe_err: 0
rx2_mpwqe_filler_cqes: 0
rx2_mpwqe_filler_strides: 0
rx2_buff_alloc_err: 0
rx2_cqe_compress_blks: 0
rx2_cqe_compress_pkts: 0
rx2_page_reuse: 0
rx2_cache_reuse: 0
rx2_cache_full: 0
rx2_cache_empty: 448
rx2_cache_busy: 0
rx2_cache_waive: 0
rx2_congst_umr: 0
rx2_arfs_err: 0
rx2_xdp_tx_xmit: 0
rx2_xdp_tx_full: 0
rx2_xdp_tx_err: 0
rx2_xdp_tx_cqes: 0
...
tx0_packets: 1
tx0_bytes: 60
tx0_tso_packets: 0
tx0_tso_bytes: 0
tx0_tso_inner_packets: 0
tx0_tso_inner_bytes: 0
tx0_csum_partial: 0
tx0_csum_partial_inner: 0
tx0_added_vlan_packets: 0
tx0_nop: 0
tx0_csum_none: 1
tx0_stopped: 0
tx0_dropped: 0
tx0_xmit_more: 0
tx0_recover: 0
tx0_cqes: 1
tx0_wake: 0
tx0_cqe_err: 0
tx1_packets: 5
tx1_bytes: 300
tx1_tso_packets: 0
tx1_tso_bytes: 0
tx1_tso_inner_packets: 0
tx1_tso_inner_bytes: 0
tx1_csum_partial: 0
tx1_csum_partial_inner: 0
tx1_added_vlan_packets: 0
tx1_nop: 0
tx1_csum_none: 5
tx1_stopped: 0
tx1_dropped: 0
tx1_xmit_more: 0
tx1_recover: 0
tx1_cqes: 5
tx1_wake: 0
tx1_cqe_err: 0
tx2_packets: 0
tx2_bytes: 0
tx2_tso_packets: 0
tx2_tso_bytes: 0
tx2_tso_inner_packets: 0
tx2_tso_inner_bytes: 0
tx2_csum_partial: 0
tx2_csum_partial_inner: 0
tx2_added_vlan_packets: 0
tx2_nop: 0
tx2_csum_none: 0
tx2_stopped: 0
tx2_dropped: 0
tx2_xmit_more: 0
tx2_recover: 0
tx2_cqes: 0
tx2_wake: 0
tx2_cqe_err: 0
...
        

Traffic Control Info

The following TC objects are supported and reported regarding the ingress filters:

• Filters

  • flower

• Actions

  • mirred

  • tunnel_key

The info is provided as one of the following events:

• Basic filter event

• Flower/IPv4 filter event

• Flower/IPv6 filter event

• Basic action event

• Mirred action event

• Tunnel_key/IPv4 action event

• Tunnel_key/IPv6 action event

General notes:

• Actions always belong to a filter, so action events share the filter event's ID via the event_id data member

• Basic filter event only contains textual kind (so users can see which real life objects' support they are lacking)

• Basic action event only contains textual kind and some basic common statistics if available

Amber Provider

Amber data for both InfiniBand and Ethernet MST devices in amBER format.

The following config files are available:

            

            
amber_devices=DEV1,DEV2,DEV3   # Default:all, or set comma separated list of devices under /dev/mst
amber_update_interval_sec=30   # Sample rate for collection amber counters
        

PPCC_ETH Provider

Programmable congestion control counters are based on an algorithm defined by an end-user, although default algorithms are also available.

Counters are collected per MST device and algorithm parameters.

The counter list depends on the installed MFT version.

A comma-separated list of device names is required to enable this provider:

            

            
ppcc_eth_devices=mt41692_pciconf0,mt41692_pciconf0.1
        

The following algorithm parameters are available:

            

            
ppcc_algo_slot=1
ppcc_algo_param_index=0
local_port=1
pnat=0
lp_msb=0
        

Fluent Aggregator

fluent_aggr listens on a port for Fluent Bit Forward protocol input connections. Received data can be streamed via a Fluent Bit exporter.

The default port is 42442. This can be changed by updating the following option:

            

            
fluent-aggr-port=42442
        

Prometheus Aggregator

prometheus_aggr polls data from a list of Prometheus endpoints.

Each endpoint is listed in the following format:

            

            
prometheus_aggr_endpoint.{N}={host_name},{host_port_url},{poll_inteval_msec}
        

Where N starts from 0.

Aggregated data can be exported via a Prometheus Aggr Exporter endpoint.

Network Interfaces

ifconfig collects network interface data. To enable, set:

            

            
enable-provider=ifconfig
        

If the Prometheus endpoint is enabled, add the following configuration to cache every collected network interface and arrange the index according to their names:

            

            
prometheus-fset-indexes=name
        

Metrices are collected for each network interface as follows:

            

            
name
rx_packets
tx_packets
rx_bytes
tx_bytes
rx_errors
tx_errors
rx_dropped
tx_dropped
multicast
collisions
rx_length_errors
rx_over_errors
rx_crc_errors
rx_frame_errors
rx_fifo_errors
rx_missed_errors
tx_aborted_errors
tx_carrier_errors
tx_fifo_errors
tx_heartbeat_errors
tx_window_errors
rx_compressed
tx_compressed
rx_nohandler
        

HCA Performance

hcaperf collects HCA performance data. Since it requires access to an RDMA device, it must use remote collection on the BlueField. On the host, the user runs the container in privileged mode and RDMA device mount.

The counter list is device dependent.

hcaperf DPU Configuration

To enable hcaperf in remote collection mode, set:

            

            
enable-provider=grpc.hcaperf
 
# specify HCAs to sample
grpc.hcaperf.mlx5_0=sample
grpc.hcaperf.mlx5_1=sample
        

hcaperf Host Configuration

To enable hcaperf in regular mode, set:

            

            
enable-provider=hcaperf
 
# specify HCAs to sample
hcaperf.mlx5_0=sample
hcaperf.mlx5_1=sample
        

NVIDIA System Management Interface

The nvidia-smi provider collects GPU and GPU process information provided by the NVIDIA system management interface.

This provider is supported only on x86_64 hosts with installed GPUs. All GPU cards supported by nvidia-smi are supported by this provider.

The counter list is GPU dependent. Additionally, per-process information is collected for the first 20 (by default) nvidia_smi_max_processes processes.

Counters can be either collected as string data "as is" in nvidia-smi or converted to numbers when nvsmi_with_numeric_fields is set.

To enable nvidia-smi provider and change parameters, set:

            

            
enable-provider=nvidia-smi
 
# Optional parameters:
#nvidia_smi_max_processes=20
#nvsmi_with_numeric_fields=1
        

NVIDIA Data Center GPU Manager

The dcgm provider collects GPU information provided by the NVIDIA data center GPU manager (DCGM) API.

This provider is supported only on x86_64 hosts with installed GPUs, and requires running the nv-hostengine service (refer to DCGM documentation for details).

DCGM counters are split into several groups by context:

• GPU – basic GPU information (always)

• COMMON – common fields that can be collected from all devices

• PROF – profiling fields

• ECC – ECC errors

• NVLINK / NVSWITCH / VGPU – fields depending on the device type

To enable DCGM provider and counter groups, set:

            

            
enable-provider=dcgm
 
dcgm_events_enable_common_fields=1
#dcgm_events_enable_prof_fields=0
#dcgm_events_enable_ecc_fields=0
#dcgm_events_enable_nvlink_fields=0
#dcgm_events_enable_nvswitch_fields=0
#dcgm_events_enable_vgpu_fields=0
        

BlueField Performance

The bfperf provider collects calculated performance counters of BlueField Arm cores. It requires the executable bfperf_pmc, which is integrated in the DOCA BFB bundle of BlueField-3, as well as an active DPE.

To enable BlueField performance provider, set:

            

            
enable-provider=bfperf
        

Diagnostic Data

Diagnostic data is comprised of two providers which gather diagnostic data counters from network interface cards (NICs). These providers support the same counters (as defined in a YAML file), but they differ in usage and collection frequency:

• Low frequency provider is defined in dts_config.ini and is controlled by DTS collection loop

• High frequency provider is defined in dts_high_freq_config.ini and operates in a distinct flow for a limited duration

The fwctl and mlx5_fwctl drivers (supported on NVIDIA networking devices from BlueField-3 and ConnectX-7 and onward) are required for firmware interaction and are part of MLNX_OFED driver. To load them, run:

            

            
modprobe -a fwctl mlx5_fwctl
        

This provider utilizes DOCA Telemetry library for data collection. Both providers get the counter set from a YAML file.

Diagnostic Data Low Frequency

To enable the diagnostic data low frequency provider, set:

            

            
enable-provider=diagnostic_data_low_freq
        

To verify that the YAML file name matches the connected NIC's type:

            

            
diagnostic-data-yml-file=/config/diagnostic_data_configs/all-single-port.yml
        

To configure the diagnostic data timestamp collection type, set the following:

            

            
diagnostic-data-timestamp-collection-type=<method>
        

Where <method> can be one of the following:

• no_counters – Do not collect timestamp counters. Default.

• start_and_end – Collect sample start and end timestamps

• per_counter – Collect every counter collection timestamp

To configure the clock firmware should use when collecting time stamps, set the following:

            

            
diagnostic-data-timestamp-source=<clock>
        

Where <clock> can be one of the following:

• RTC - Real-time clock. Default.

• RFC - Free-running clock

Diagnostic Data High Frequency

The Diagnostic Data High Frequency provider supports higher sampling frequencies with sub-millisecond resolution, enabling detailed and precise telemetry collection. Due to the large volume of collected data, this provider is designed to run ad-hoc for limited periods, unlike standard DTS providers configured via the DTS configuration file at /opt/mellanox/doca/services/telemetry/config/dts_config.ini.

While the standard DTS flow functions as an endless collect-export loop, High Frequency Telemetry (HFT) operates as an external flow triggered by the HFT configuration file, located at /opt/mellanox/doca/services/telemetry/config/clx_ad_hoc_runner_config.ini .

This configuration file specifies the HFT session's timing parameters, provider settings, and export configurations. HFT allows data export to endpoints or protocols distinct from those used in the standard DTS collection loop. The DTS configuration references the HFT configuration file as an "ad hoc runner," enabling DTS to monitor its status via the following directive:

            

            
ad-hoc-runner-file=/config/clx_ad_hoc_runner_config.ini
        

The HFT configuration file serves as both the trigger and definition for HFT sessions. Modifying the file stops the current session and applies the new configuration for the next session. Deleting the file stops any pending HFT sessions entirely.

Required HFT Parameters

This table provides the details of the required HFT parameters. Refer to section "HFT Configuration File Example" for more helpful tips.

Provider Compatibility

Both low and high frequency providers can run concurrently. The low frequency provider samples at the DTS standard frequency (defined in dts_config.ini), and the high frequency provider samples counters based on the HFT configuration file ( dts_high_freq_config.ini ).

To allow both providers to run concurrently, verify that the counters, the timestamp collection type, and the timestamp collection source are identical. Otherwise, when the high frequency provider starts sampling, the low frequency provider hangs until the end of the HFT session.

HFT Configuration File Example

            

            
## DTS configuration file for ad-hoc high frequency collection
## When modified, the file is parsed and applied.
## Note that the folders path is the container path, not the host path.
 
## Each section defines a collection. A file may have several sections, each one defines a high frequency collection.
## Section names must be unique and will be used as collection name by clx.
[hft-collection-session]
 
### Time between samples in microseconds
sample-time-us=100000
 
### Start time of high frequency collection. Can be in the format HH:MM:SS or HH:MM or as epoch timestamp in microseconds
### Note - in container, the time is in UTC
start-time=18:00:00
 
### End time of high frequency collection. Can be in the format HH:MM:SS or HH:MM or as epoch timestamp in microseconds
### Note - in container, the time is in UTC
end-time=18:01:00
### Alternatively, you can set the number of iterations. This and start_time field will determine the end time
#num-iterations=300
 
### Data provider to use
provider=diagnostic_data_high_freq
 
### Write data to file system. Could potentially fill up the disk
file-write=false
 
### Root directory to store the data
# Ignored if file-write is set to false
data-root=/data
 
### Enable busy wait between iterations, for a more accurate sample time (default is false)
#busy-wait-sampling=true
 
### Set prometheus endpoint to enable http endpoint
#prometheus-endpoint=http://0.0.0.0:9112
 
### Set fluentbit config dir to enable fluentbit export
#fluentbit-config-dir=/config/fluent_bit_configs
 
### Set open telemetry receiver to enable open telemetry export
#open-telemetry-receiver=http://0.0.0.0:9502/v1/metrics
 
### Set remote write receiver to enable remote write export
#remote-write-receiver=http://0.0.0.0:9090/api/v1/write
 
### Provider specific parameters. Format is 'provider.$KEY=$VALUE'.
### The options below are specific to the diagnostic data high frequency provider
 
# Number of samples to collect on each iteration
provider.diagnostic-data-num-samples=1000
 
# The time period (in nanoseconds) between samples
provider.diagnostic-data-sample-period-nsec=100000
 
# The YAML file with the configuration for the diagnostic data provider
provider.diagnostic-data-yml-file=/config/diagnostic_data_configs/all-single-port.yml
 
# Diagnostic Data timestamp collection type. Options are ['no_counters', 'start_and_end', 'per_counter']. default: 'no_counters'
#provider.diagnostic-data-timestamp-collection-type=start_and_end
 
# Diagnostic Data timestamp source. Options are ['RTC', 'FRC']. default: 'RTC'
#provider.diagnostic-data-timestamp-source=FRC
        

Diagnostic Data YAML File

For compatibility with other related tools, the counter set is defined in YAML format.

There are 4 YAML files within a DTS container (one per permutation of BlueField-3 and ConnectX-7 with dual or single ports). The path to the YAMLs folder is /opt/mellanox/doca/services/telemetry/config/diagnostic_data_configs which is mounted to /config/diagnostic_data_configs.

By default, YAML files include a counter set that is not device-specific. This implies that the same counter set is utilized across all devices by default.

It is possible to assign a specific device within a YAML file; however, this requires maintaining a separate copy of the YAML file for each device. To manage multiple devices, use the diagnostic-data-yml-dir option to specify a directory for YAML files, where each .yml/.yaml file is utilized. This folder should be available to the container under /opt/mellanox/doca/services/telemetry/config .

The following list describes the expected entries in the YAML file:

• counters – sequence of counters to collect

  • id – counter data ID

  • desc – counter description (optional)

  • unit – name of unit to collect from (optional)

  • name – name of counter to use (optional). If not specified, the generated name is based on the counter description. Otherwise, it is based on the data ID.

• device – name of the mlx device to collect (optional). If not used, the provider requires a single file containing a list of counters, which it then applies to all available devices on the host.

YAML File Example

The following is the content of the all-dual-port.yml file provided by DTS:

            

            
counters:
  - id: 0x1020000100000000
    name: port_rx_bytes_0
  - id: 0x1020000100000001
    name: port_rx_bytes_1
  - id: 0x1020000300000000
    name: port_rx_packets_0
  - id: 0x1020000300000001
    name: port_rx_packets_1
  - id: 0x1140000100000000
    name: port_tx_bytes_0
  - id: 0x1140000100000001
    name: port_tx_bytes_1
  - id: 0x1140000300000000
    name: port_tx_packets_0
  - id: 0x1140000300000001
    name: port_tx_packets_1
  - id: 0x1100000100000000
    name: port_tx_transport_cnp_sent_packets_0
  - id: 0x1100000100000001
    name: port_tx_transport_cnp_sent_packets_1
  - id: 0x1080000500000000
    name: port_rx_transport_cnp_handled_packets_0
  - id: 0x1080000500000001
    name: port_rx_transport_cnp_handled_packets_1
  - id: 0x1080000400000000
    name: port_rx_transport_ecn_packets_0
  - id: 0x1080000400000001
    name: port_rx_transport_ecn_packets_1
  - id: 0x1160000b00000000
    name: pcie_link_latency_total_read_ns
  - id: 0x1160000c00000000
    name: pcie_link_latency_total_read_packets
  - id: 0x1160000d00000000
    name: pcie_link_latency_max_read_ns
  - id: 0x1160000e00000000
    name: pcie_link_latency_min_read_ns
  - id: 0x1180000100000000
    name: global_icmc_request
  - id: 0x1180000200000000
    name: global_icmc_hit
  - id: 0x1180000300000000
    name: global_icmc_miss
        

RDMA Notifications

The RDMA Notifications events provider collects notifications from the RDMA firmware and converts them into DTS events for telemetry purposes.

To enable the RDMA notifications provider, add the following line to the DTS configuration file (/opt/mellanox/doca/services/telemetry/config/dts_config.ini):

            

            
enable-provider=rdma_notifications
        

Prerequisites

• RDMA device(s) configured to operate in Ethernet protocol

• OFED installed

• ConnectX-7 or BlueField-3 and later

• Firmware version x.44.0820 or newer

Options

Events List

There are two types of events provided by the RDMA notifications:

• Notification Event – A direct translation of an RDMA notification into one of four possible event types, based on the notification content

• Counter Event – A periodic aggregation event that reports the total number of events collected for each syndrome

Notification Events

Additional fields are provided for convenience:

• guid – A globally unique identifier for the device

• idx – The event index, determined by the provider option rdma-notifications-num-event-indexes

• device_name – The name of the RDMA device, based on the devices listed in rdma-notifications-hca

These fields are not specified in the NVIDIA Adapters PRM but are included in the event data to enhance usability.

cqe_with_error_responder_ipv4

            

            
telemetry_counter
msn
syndrome
qp_type
vport_id
source_qpn
destination_qpn
psn
timestamp
dest_ipv4
node_guid
idx
device_name
        

cqe_with_error_responder_ipv6

            

            
telemetry_counter
msn
syndrome
qp_type
vport_id
source_qpn
destination_qpn
psn
timestamp
dest_ipv6
node_guid
idx
device_name
        

cqe_with_error_requestor_ipv4

            

            
telemetry_counter
msn
syndrome
opcode
qp_type
vport_id
source_qpn
destination_qpn
psn
timestamp
dest_ipv4
node_guid
idx
device_name
        

cqe_with_error_requestor_ipv6

            

            
telemetry_counter
msn
syndrome
opcode
qp_type
vport_id
source_qpn
destination_qpn
psn
timestamp
dest_ipv6
node_guid
idx
device_name
        

Counter Events

Describing the number of events per syndrome, for each device.

cqe_with_error_syndrome_counter

            

            
syndrome
device_name
node_guid
CQEwE_events
        

Data Outputs

DTS can send the collected data to the following outputs:

• Data writer (saves binary data to disk)

• Prometheus endpoint (keeps the most recent data to be pulled)

• Fluent Bit (push-model streaming)

• Open Telemetry exporter (push-model streaming)

• Prometheus Remote write exporter (push-model streaming)

Data Writer

The data writer is disabled by default to save space on BlueField. Steps for activating data write during debug can be found under section Enabling Data Write.

The schema folder contains JSON-formatted metadata files which allow reading the binary files containing the actual data. The binary files are written according to the naming convention shown in the following example (apt install tree):

            

            
tree /opt/mellanox/doca/services/telemetry/data/
/opt/mellanox/doca/services/telemetry/data/
├── {year}
│   └── {mmdd}
│        └── {hash}
│             ├── {source_id}
│             │   └── {source_tag}{timestamp}.bin
│             └── {another_source_id}
│                  └── {another_source_tag}{timestamp}.bin
└── schema
    └── schema_{MD5_digest}.json
        

New binary files appears when the service starts or when binary file age/size restriction is reached. If no schema or no data folders are present, refer to the Troubleshooting section.

Reading the binary data can be done from within the DTS container using the following command:

            

            
crictl exec -it <Container ID> /opt/mellanox/collectx/bin/clx_read -s /data/schema /data/path/to/datafile.bin
        

Example output:

            

            
{
    "timestamp": 1634815738799728,
    "event_number": 0,
    "iter_num": 0,
    "string_number": 0,
    "example_string": "example_str_1"
}
{
    "timestamp": 1634815738799768,
    "event_number": 1,
    "iter_num": 0,
    "string_number": 1,
    "example_string": "example_str_2"
}
…
        

Prometheus

The Prometheus endpoint keeps the most recent data to be pulled by the Prometheus server and is enabled by default.

To check that data is available, run the following command on BlueField:

            

            
curl -s http://0.0.0.0:9100/metrics
        

The command dumps every counter in the following format:

            

            
counter_name {list of meta fields} counter_value timestamp
        

Additionally, endpoint supports JSON and CSV formats:

            

            
curl -s http://0.0.0.0:9100/json/metrics
curl -s http://0.0.0.0:9100/csv/metrics
        

Configuration Details

Prometheus is configured as a part of dts_config.ini.

By default, the Prometheus HTTP endpoint is set to port 9100. Comment this line out to disable Prometheus export.

            

            
prometheus=http://0.0.0.0:9100
        

Prometheus can use the data field as an index to keep several data records with different index values. Index fields are added to Prometheus labels.

            

            
# Comma-separated counter set description for Prometheus indexing:
#prometheus-indexes=idx1,idx2
 
# Comma-separated fieldset description for prometheus indexing
#prometheus-fset-indexes=idx1,idx2
        

The default fset index is device_name. It allows Prometheus to keep ethtool data up for both the p0 and p1 devices.

            

            
prometheus-fset-indexes=device_name
        

If fset index is not set, the data from p1 overwrites p0's data.

For quick name filtering, the Prometheus exporter supports being provided with a comma-separated list of counter names to be ignored:

            

            
#prometheus-ignore-names=counter_name1,counter_name_2
        

For quick filtering of data by tag, the Prometheus exporter supports being provided with a comma-separated list of data source tags to be ignored.

Users should add tags for all streaming data since the Prometheus exporter cannot be used for streaming. By default, FI_metrics are disabled.

            

            
prometheus-ignore-tags=FI_metrics
        

Prometheus Aggregator Exporter

Prometheus aggregator exporter is an endpoint that keeps the latest aggregated data using prometheus_aggr.

This exporter labels data according to its source.

To enable this provider, users must set 2 parameters in dts_config.ini:

            

            
prometheus-aggr-exporter-host=0.0.0.0
prometheus-aggr-exporter-port=33333
        

Fluent Bit

Fluent Bit allows streaming to multiple destinations. Destinations are configured in .exp files that are documented in-place and can be found under:

            

            
/opt/mellanox/doca/services/telemetry/config/fluent_bit_configs
        

Fluent Bit allows exporting data via "Forward" protocol which connects to the Fluent Bit/FluentD instance on customer side.

Export can be enabled manually:

1. Uncomment the line with fluent_bit_configs=… in dts_config.ini.

2. Set enable=1 in required .exp files for the desired plugins.

3. Additional configurations can be set according to instructions in the .exp file if needed.

4. Restart the DTS.

5. Set up receiving instance of Fluent Bit/FluentD if needed.

6. See the data on the receiving side.

Export file destinations are set by configuring .exp files or creating new ones. It is recommended to start by going over documented example files. Documented examples exist for the following supported plugins:

• forward

• file

• stdout

• kafka

• es (elastic search)

• influx

Export File Configuration Details

Each export destination has the following fields:

• name – configuration name

• plugin_name – Fluent Bit plugin name

• enable – 1 or 0 values to enable/disable this destination

• host – the host for Fluent Bit plugin

• port – port for Fluent Bit plugin

• msgpack_data_layout – the msgpacked data format. Default is flb_std. The other option is custom. See section Msgpack Data Layout for details.

• plugin_key=val – key-value pairs of Fluent Bit plugin parameter (optional)

• counterset/fieldset – file paths (optional). See details in section Cset/Fset Filtering.

• source_tag=source_tag1,source_tag2 – comma-separated list of data page source tags for filtering. The rest tags are filtered out during export. Event tags are event provider names. All counters can be enabled/disabled only simultaneously with a counters keyword.

Msgpack Data Layout

Data layout can be configured using .exp files by setting msgpack_data_layout=layout. There are two available layouts: Standard and Custom.

The standard flb_std data layout is an array of 2 fields:

• timestamp double value

• a plain dictionary (key-value pairs)

The standard layout is appropriate for all Fluent Bit plugins. For example:

            

            
[timestamp_val, {"timestamp"->ts_val, type=>"counters/events", "source"=>"source_val", "key_1"=>val_1, "key_2"=>val_2,...}]
        

The custom data layout is a dictionary of meta-fields and counter fields. Values are placed into a separate plain dictionary. Custom data format can be dumped with stdout_raw output plugin of Fluent-Bit installed or can be forwarded with forward output plugin.

Counters example:

            

            
{"timestamp"=>timestamp_val, "type"=>"counters", "source"=>"source_val", "values"=> {"key_1"=>val_1, "key_2"=>val_2,...}}
        

Events example:

            

            
{"timestamp"=>timestamp_val, "type"=>"events", "type_name"=>"type_name_val", "source"=>" source_val", "values"=>{"key_1"=>val_1, "key_2"=>val_2,...}}
        

Cset/Fset Filtering

Each export file can optionally use one cset and one fset file to filter DTS counters and events data.

• cset contains tokens per line to filter data with "type"="counters".

• fset contains several blocks started with the header line [event_type_name] and tokens under that header. An Fset file is used to filter data with "type"="events".

  Note

  Event type names could be prefixed to apply the same tokens to all fitting types. For example, to filter all ethtool events, use [ethtool_event_*].

If several tokens must be matched simultaneously, use <tok1>+<tok2>+<tok3>. Exclusive tokens are available as well. For example, the line <tok1>+<tok2>-<tok3>-<tok4> filters names that match both tok1 and tok2 and do not match tok3 or tok4.

The following are the details of writing cset files:

            

            
# Put tokens on separate lines
# Tokens are the actual name 'fragments' to be matched
# port$ # match names ending with token "port"
# ^port # match names starting with token "port"
# ^port$ # include name that is exact token "port
# port+xmit # match names that contain both tokens "port" and "xmit"
# port-support # match names that contain the token "port" and do not match the "-" token "support"
#
# Tip: To disable counter export put a single token line that fits nothing
        

The following are the details of writing fset files:

            

            
# Put your events here
# Usage:
#
# [type_name_1]
# tokens
# [type_name_2]
# tokens
# [type_name_3]
# tokens
# ...
# Tokens are the actual name 'fragments' to be matched
# port$ # match names ending with token "port"
# ^port # match names starting with token "port"
# ^port$ # include name that is exact token "port
# port+xmit # match names that contain both tokens "port" and "xmit"
# port-support # match names that contain the token "port" and do not match the "-" token "support"
 
# The next example will export all the "tc" events and all events with type prefix "ethtool_" "ethtool" are filtered with token "port":
# [tc]
#
# [ethtool_*]
# packet
 
# To know which event type names are available check export and find field "type_name"=>"ethtool_event_p0"
# ...
# Corner cases:
# 1. Empty fset file will export all events.
# 2. Tokens written above/without [event_type] will be ignored.
# 3. If cannot open fset file, warning will be printed, all event types will be exported.
        

NetFlow Exporter

NetFlow exporter must be used when data is collected as NetFlow packets from the telemetry client applications. In this case, DOCA Telemetry Exporter NetFlow API sends NetFlow data packages to DTS via IPC. DTS uses NetFlow exporter to send data to the NetFlow collector (3rd party service).

To enable NetFlow exporter, set netflow-collector-ip and netflow-collector-port in dts_config.ini. netflow-collector-ip could be set either to IP or an address.

For additional information, refer to the dts_config.ini file.

Open Telemetry Exporter

DTS is capable of streaming telemetry data towards an Open Telemetry receiver using the HTTP-based OTLP metrics protocol.

The Open Telemetry exporter is disabled by default and can be configured via dts_config.ini.

To enable the telemetry data streaming in the form of metrics, set the destination of the OTLP data streaming (i.e., the Open Telemetry receiver):

            

            
open-telemetry-receiver=http://0.0.0.0:9502/v1/metrics
        

This example expects the running Open Telemetry receiver to be bound to port 9502 on the host running the DTS instance. Section "Open Telemetry Metrics Receiver Sample Configuration" provides an example of the Open Telemetry receiver configuration.

The Open Telemetry exporter can be configured to apply counter-set or field-set to the data stream, similar to the Fluent Bit exporter:

            

            
open-telemetry-counter-set=slow_counters
open-telemetry-field-set=slow_counters
        

The Open Telemetry exporter sends data in bulk, combining bursts when possible. By default, the size of the bulk is up to 100 data points. If desirable, the size of the bulk can be altered using the open-telemetry-bulk-size configuration parameter. For example:

            

            
open-telemetry-bulk-size=20
        

Open Telemetry Metrics Receiver Sample Configuration

The following is a sample Open Telemetry server configuration file (collector-config.yaml):

            

            
receivers:
  otlp:
    protocols:
      http:
        endpoint: 0.0.0.0:9502
 
exporters:
  debug:
    verbosity: detailed
 
service:
  pipelines:
    metrics:
      receivers: [otlp]
      exporters: [debug]
        

To run an Open Telemetry instance:

            

            
docker run --rm --name "$(whoami)_open_telemetry" -p 9502:9502 \
    -v $(pwd)/data:/data -v $(pwd)/collector-config.yaml:/etc/otelcol/config.yaml otel/opentelemetry-collector
        

Prometheus Remote Write Exporter

DTS can stream the telemetry data towards an external Prometheus service with an enabled metrics receiver. If enabled, the DTS data exporter acts as a Prometheus remote write protocol "Sender", streaming telemetry data towards a Prometheus server using remote write protocol.

To enable telemetry data streaming:

            

            
remote-write-receiver=http://0.0.0.0:9090/api/v1/write
        

This example assumes the running Prometheus remote write receiver to be bound to port 9090 on the host running the DTS instance.

For example, a Prometheus remote write receiver could be run using the following command:

            

            
docker run -p 9090:9090 prom/prometheus --config.file=/etc/prometheus/prometheus.yml \
    --storage.tsdb.path=/prometheus --web.console.libraries=/usr/share/prometheus/console_libraries \
    --web.console.templates=/usr/share/prometheus/consoles --web.enable-remote-write-receiver
        

The Prometheus remote write exporter can be configured to apply counter-set or field-set to the data stream, similar to the Open Telemetry exporter:

            

            
remote-write-counter-set=slow_counters
remote-write-field-set=slow_counters
        

The Prometheus remote write exporter sends data in bulk, combining bursts when possible. By default, the size of the bulk is up to 100 data points. If necessary, the size of the bulk can be altered using the bulk-size configuration parameter. For example:

            

            
remote-write-bulk-size=20
        

Load Balancer Exporter

The load balancer exporter is an exporter which distributes or copies data between multiple instances of another (secondary) exporter.

The load balancer exporter plugin is customized using configuration variables which may come from config file, environment, or command line.

For example, the number of instances may be set as follows depending on the use case:

• Line in a configuration file – loadbalancer-num-instances=5

• Command line option – -X loadbalancer-num-instances=5

• Environment variable – CLX_LOADBALANCER_NUM_INSTANCES=5

  Info

  In this instance, dashes (-) in variable names are substituted with underscores (_) and the CLX_ prefix is added.

loadbalancer-exporter-config-var is the name of another configuration variable, which is substituted for all secondary exporter instances. For example:

• If loadbalancer-exporter-config-var=open-telemetry-receiver and loadbalancer-num-instances=3, then instance configuration variables open-telemetry-receiver-0, open-telemetry-receiver-1, and open-telemetry-receiver-2 should exist, and the value of the open-telemetry-receiver configuration variable is substituted to the value of open-telemetry-receiver-0 for the first instance of the secondary exporter, open-telemetry-receiver-1 for the second instance, etc.

• If loadbalancer-exporter-config-var does not exist, all instances are created equal, and their configuration is done according to other available configuration. If any of the instance configuration variables do not exist, that instance is created with other available configuration.

The following is a section of the dts_config.ini configuration file which provides a sample of enabling load balancer to run 2 OTLP exporter instances:

            

            
loadbalancer-exporter-name=open_telemetry
loadbalancer-num-instances=2
loadbalancer-exporter-config-var=open-telemetry-receiver
loadbalancer-mode=replicate
open-telemetry-receiver-0=http://10.141.160.210:9502/v1/metrics
open-telemetry-receiver-1=http://10.141.160.211:9502/v1/metrics
        

The DOCA Privileged Executer (DPE) is a daemon that enables specific DOCA services, including DTS, to access BlueField information that would otherwise be inaccessible from a container due to technology limitations or restricted permissions. By granting access to privileged system information, DPE enriches the data collected by DTS, providing additional insights and metrics. However, DTS can still operate independently without DPE, as the daemon is disabled by default.

DPE Usage

DPE is controlled by systemd, and can be used as follows:

• To check DPE status:

  Copy

  Copied!

              
  
              
  sudo systemctl status dpe
          

• To start DPE:

  Copy

  Copied!

              
  
              
  sudo systemctl start dpe
          

• To stop DPE:

  Copy

  Copied!

              
  
              
  sudo systemctl stop dpe
          

DPE logs can be found in /var/log/doca/telemetry/dpe.log.

DPE Configuration File

DPE can be configured by the user. This section covers the syntax and implications of its configuration file.

The DPE configuration file allows users to define the set of commands that DPE should support. This may be done by passing the -f option in the following line of /etc/systemd/system/dpe.service:

            

            
ExecStart=/opt/mellanox/doca/services/telemetry/dpe/bin/dpeserver -vvv
        

To use the configuration file:

            

            
ExecStart=/opt/mellanox/doca/services/telemetry/dpe/bin/dpeserver -vvv -f /path/to/dpe_config.ini
        

The configuration file supports the following sections:

• [server] - list of key=value lines for general server configuration. Allowed keys: socket.

• [commands] - list of bash command lines that are not using custom RegEx

• [commands_regex] - list of bash command lines that are using custom RegEx

• [regex_macros] - custom RegEx definitions used in the commands_regex section

Consider the following example configuration file:

            

            
[server]
socket=/tmp/dpe.sock
 
[commands]
hostname
cat /etc/os-release
 
[commands_regex]
crictl inspect $HEXA       # resolved as "crictl inspect [a-f0-9]+"
lspci $BDF                 # resolved as "lspci ([0-9a-f]{4}\:|)[0-9a-f]{2}\:[0-9a-f]{2}\.[0-9a-f]"
 
[regex_macros]
HEXA=[a-f0-9]+
BDF=([0-9a-f]{4}\:|)[0-9a-f]{2}\:[0-9a-f]{2}\.[0-9a-f]
        

This chapter provides an overview and deployment configuration of DOCA Telemetry Service with Grafana .

Grafana Deployment Prerequisites

• BlueField DPU running DOCA Telemetry Service.

• Optional remote server to host Grafana and Prometheus.

• Prometheus installed on the host machine. Please refer to the Prometheus website for more information.

• Grafana installed on the host machine. Please refer to Grafana Labs website for more information.

Grafana Deployment Configuration

DTS Configuration (DPU Side)

Configuring DTS to export the sysfs counter using the Prometheus plugin:

1. Make sure the sysfs counter is enabled.

   Copy

   Copied!

               
   
               
   vim /opt/mellanox/doca/services/telemetry/config/dts_config.ini
    
   enable-provider=sysfs
           

2. Enable Prometheus exporter by setting the prometheus address and port.

   Copy

   Copied!

               
   
               
   vim /opt/mellanox/doca/services/telemetry/config/dts_config.ini
    
   prometheus=http://0.0.0.0:9100
           

   Note

   In this example, the Prometheus plugin exports data on localhost port 9100, this is an arbitrary value and can changed.

Prometheus Configuration (Remote Server)

Please download Prometheus for your platform.

Prometheus is configured via command-line flags and a configuration file, prometheus.yml.

1. Open the prometheus.yml file and configure the DPU as the endpoint target.

   Copy

   Copied!

               
   
               
   vim prometheus.yml
   # metrics_path defaults to '/metrics'
   # scheme defaults to 'http'.
    
   static_configs:
   - targets: ["<dpu-ip>:<prometheus-port>"]
           

   Where:

   • <dpu-ip> is the DPU IP address. Prometheus reaches to this IP to pull data.

   • <prometheus-port> the exporter port that set in DTS configuration.

2. Run Prometheus server:

   Copy

   Copied!

               
   
               
   ./prometheus --config.file="prometheus.yml"
           

   Tip

   Prometheus services are available as Docker images. Please refer to Using Docker in Prometheus' Installation guide.

Grafana Configuration (Remote Server)

Please download and install Grafana for your platform.

1. Setup Grafana. Please refer to Install Grafana guide in Grafana documentation.

2. Log into the Grafana dashboard at http://localhost:3000.

   Note

   Port 3000 is the default port number set by Grafana. This can be changed if needed. The default credentials are admin/admin.

3. Add Prometheus as data source by navigating to Settings → Data sources → Add data source → Prometheus.

   

4. Configure the Prometheus data source. Under the HTTP section, set the Prometheus server address.

   

   Note

   The Prometheus server's default listen port is 9090. Prometheus and Grafana are both running on the same server, thus the address is localhost.

5. Save and test.

Exploring Telemetry Data

Go to the Explore page on the left-hand side, and choose a Prometheus provider.

Choose a metric to display and specify a label. The label can be used to filter out data based on the source and HCA devices.

Graph display after selecting a metric and specifying a label to filter by:

On top of the Troubleshooting section in the NVIDIA DOCA Container Deployment Guide, here are additional troubleshooting tips for DTS:

• For general troubleshooting, refer to the DOCA Troubleshooting.

• If the pod's state fails to be marked as "Ready", refer to /var/log/syslog.

• Check if the service is configured to write data to the disk as this may cause the system to run out of disk space.

• If a PIC bus error occurs, configure the following files inside the container:

  Copy

  Copied!

              
  
              
  crictl exec -it <container-id> /bin/bash
  # Add to /config/clx.env the following line:
  "
  export UCX_TLS=tcp
  "
          

• If a wrong hostname is set in the source/source_id label of DTS exports, this may indicate that the container is not running on the host network. In such cases, usually the hostname is based on the container name doca-telemetry. To fix this, update the docker command/YAML file accordingly.