Client API Reference, v2.2.5

Return the state

A message containing relevant information for the abort.

List of collective implementations

The infrastructure profile to use for the binding.

previous binding (if any), useful for incremental changes to existing bindings

Assigns platforms to different regions of the infrastructure. DSE server will raise an error if this is non-empty and feature flag onearm is not being used.

The platform type. Typically obtained from CreateBinding()

The created binding object

result ids can be obtained from the trial_report, provided by ListTrialReports api call If result_ids is empty, then it will collect diagnostics for the currently configured trial list of result ids to include in the diagnostic file

filepath to the archived log files (zip, tar, etc.)

URL to download the diagnostic file

Sequence number of the log message.

The timestamp signifies the time at which the log message was generated.

The severity level of the log message.

The name of the component emitting the log message.

The workspace that the trial will be stored under. If the workspace does not exist, it will be created.

A list of tags associated with this trial.

The type of platform to run the workload over

configuration settings specific to nccl

TCP transport configuration

RoCEv2 transport configuration

BEGIN PRIVATE Falcon transport configuration

KCCB configuration

Workload replay configuration

typically obtained from CreateBinding()

hold version info

List of impairment metrics

The unique identifier for the trial

The is_readonly is set if the config cannot be used to run a trial. E.g. part of the config not exposed by the UI is removed before saving. The configuration can be inspected in the UI, but a trial will fail to run.

The timestamp signifies the time at which a trial run was started. In storage, each trial directory is named according to the following format: ISO 8601 formatted timestamp YYYY-MM-DDTHH:mm:SS.ms:timezone

workspace name

The storage path of the trial directory.

Tags associated with the trial.

Description of the trial

Stores the current state of this trial.

System tags associated with the trial.

The timestamp signifies the time at which a trial run was completed.

A table of nccl-like summary results

A table of workload replay summary results

A collection of artifacts generated by the trial that are saved in storage.

A collection of artifacts generated by the trial that are saved in storage.

A flag to indicate whether to use the v2 report format

Detailed report information in v2 format

Workspace name

List of tags to allow filtering trial results

A system supplied collective algorithm

A path of format package.module.classname to a class that inherits from the keys_ai_ml_chakra.Expander. See the EXPANDERS.md in the keys_ai_ml_chakra package for details on how to create your own custom expander class.

Configuration for flow control mechanism used by this algorithm.

Chassis IP Address or FQDN

Chassis port. Formats <front-panel-port> or <front-panel-port>.<fanout>. Examples: '1' or '1.4'

The type of collective communication operation.

A system-provided or user-provided algorithm specification used to expand a collective communication operation into a set of peer-to-peer operations.

ECN configuration

PFC configuration

DCQCN rate control configuration

Enable/disable DCQCN

Factor to update Alpha every update period (fixed point fraction of 2^10)

Alpha update period (microseconds)

Initial Alpha value

Current and target rate limit set after first CNP (Mbps)

Maximal ratio of rate decrease in a single event (percentage)

Minimal rate limit of the QP (Mbps)

The coefficient between Alpha and the rate reduction factor

The minimum time period between rate reductions (microseconds)

If enabled, whenever a CNP is processed, the target rate will be updated to the current rate

The time period between rate increase events (microseconds)

The sent bytes counter between rate increase events (64B)

The threshold of rate increase events for moving to next rate increase phase

The rate increase value in the Additive Increase phase (Mbps)

The rate increase value in the Hyper Increase phase (Mbps)

Minimal time between two consecutive CNPs sent (microseconds)

The message to echo

The echoed message

DSCP of CNP packets

Configures the ECN bits for data packets

Configures the ECN bits for control packets; eg RoCEv2 ACKs

Configures the ECN bits for CNP packets

(Maximum) RMDA message size in Bytes

Number of Queue Pairs per rank pair

Queue Pair Negotiation method

RDMA verb to use for data transfers in collective communication operations

TCPStore hostname or IP address (only applicable when qp_negotiation is METHOD_TCP_STORE)

TCPStore port number (only applicable when qp_negotiation is METHOD_TCP_STORE)

Enable flow control mechanism where receiver grants credits to sender to control the data flow.

Maximum number of credits allowed in flight.

Delay between data reception and credit transmission (microseconds)

Determines how credits are distributed across streams. One credit can either unblock one stream at a time or all streams simultaneously

IP address

IP prefix

Gateway IP address

IP address

IP prefix

Gateway IP address

Custom Ixperf environment variables

Speed, modulation and FEC mode

Enable/disable auto negotiation

Enable/disable link training

Enable/disable IEEE Defaults. This setting takes precedence over auto-negotiation and link training

Custom NCCL environment variables

Ethernet Maximum transmission unit

IPv4 Addressing specifics

IPv6 Addressing specifics

BEGIN PRIVATE Quality of service

END PRIVATE Congestion control

BEGIN PRIVATE Packet capture END PRIVATE

MAC address

VLAN Tags

ROCEv2 Transport specific settings

Enable/disable packet capture

Size of the capture file size in bytes.

What to do when the capture buffer is full

Size of each packet slice in bytes. If set to 0, the entire packet is captured

Enable/disable Priority Flow Control

List of priorities

Pirority Trust mode

DSCP to priority mapping

Priority to traffic class mapping

(Maximum) RMDA message size in Bytes

Number of Queue Pairs per rank pair

Queue Pair Negotiation method

RDMA verb to use for data transfers in collective communication operations

TCPStore hostname or IP address (only applicable when qp_negotiation is METHOD_TCP_STORE)

TCPStore port number (only applicable when qp_negotiation is METHOD_TCP_STORE)

The AckTimeout for RoCEv2 protocol

The RetransRetryCount for RoCEv2 protocol

Triggers retransmission when RNR NACK is received

Request ACK after every N packets

Enable/disable reuse of Queue Pairs (QPs) for RDMA messages between a rank pair

Enable/disable the Rx Reordering

Configures ACK DSCP value for RoCEv2 protocol

Configures NACK DSCP value for RoCEv2 protocol

Configures DSCP for all data traffic

Server IP address or FQDN

Server NIC interface

Enable/Disable VLAN

Currently one vlan tag is required.

VLAN Priority

VLAN ID

Rows of summary data, one per collective operation

Schema defining columns in the summary table

Paths to configuration files used in the trial

Paths to summary data files

Paths to port-level metrics files

Paths to fabric-level metrics files

Paths to packet capture files

Paths to custom user-generated files

Paths to error message log files

Paths to flow-level metrics files

Paths to data chunk metrics files

Paths to impairment metrics files

Paths to queue pair metrics files

Paths to data size breakdown metrics files

Paths to per-iteration metrics files

Paths to emulation summary metrics files

Paths to workload detail files

Paths to packet drop metrics files

Paths to packet reorder metrics files

The name identifier for this blackbox fabric instance. fabric name

The number of hosts in the fabric. This value is used to determine the number of racks and switches needed in the fabric based on the tier configurations.

Gap in field numbers lets us add this later without breaking compatibility. uint32 host_nic_radix = 6;

The maximum number of ports from each host up to its peer switch.

The schema for the host wiring.

The schema for the spine wiring.

The configurations for each tier in the fabric. The first element in the list corresponds to the leaf tier, the second to the tier above it, and so on.

non-detailed generic fabric (a blackbox)

Multi-stage Clos fabric

BEGIN PRIVATE EXPERIMENTAL: network fabric featuring multiple planes within a rack

The name of the generic host.

The number of Neural Processing Units (NPUs) in the host. N.B. generic hosts have a 1:1 ratio of nics and gpus

The custom bandwidth for the generic host in Gbps.

BEGIN PRIVATE The bandwidth of NVLink by version.

The number of hosts in the fabric.

Zionex host type.

Generic host type.

EXPERIMENTAL

END PRIVATE Whether to use NPU interconnect for the host.

A custom bandwidth for the infrastructure in Gbps.

The speed of the infrastructure fabric.

The type of host used in the infrastructure.

Specifies thes fabric topology used in the infrastructure.

The maximum number of ports under a device.

Whether to use half the maximum ports.

The maximum number of ports from each host up to its peer switch.

Whether to use the maximum number of ports.

The oversubscription ratio for downlink traffic.

The oversubscription ratio for uplink traffic.

The number of racks in the fabric.

The number of hosts per rack.

Ethernet speed for inter-rack communication via scale-out switches.

Ethernet speed for intra-rack communication on the scale-up network.

The number of NPUs in the host.

The number of NICs for scale-up traffic.

The number of NICs for scale-out traffic.

The name identifier for this RackPlane host instance.

The radix of the switch in the fabric.

The oversubscription ratio for this tier, expressed as downlink_factor:uplink_factor. This determines the ratio of downward-facing to upward-facing port bandwidth.

The port speed (bandwidth) for all switch ports on this tier. Can be specified as a custom value in Gbps or as a standard fabric speed.

The maximum number of ports that can connect from a switch on this tier to a single switch on the next higher tier. Controls link aggregation between tiers.

The maximum number of ports on each switch that can connect downward to the next lower tier. Typically uses half the switch radix by default.

The name identifier for this Zionex host instance.

communication collective algorithm

A data size iterator with start, step, and end values

A list of data size definitions

Number of benchmark runs per data size

The channels topology configuration specifying how ranks are distributed and connected.

The delay (in ms) between repeated executions of the benchmark.

The benchmark configuration defining the collective communication test.

The initial data size of a benchmark collective operation.

The amount to increment the data size between iterations

maximum data size after which the benchmark completes

the size of data in bytes

Custom binding configuration that maps ranks to NPUs, configures NIC settings, and assigns physical test resources.

the Binding keeps a copy of the infrastructure profile so that Trials can be re-run with the original versions of (potentially modified) profiles.InfraProfile

low-level chakra infrastructure matching infrastructure_profile

Annotations for low-level chakra infrastructure.

Assigns platforms to different regions of the infrastructure. Should be populated by the server and be a copy of the platform_regions sent in CreateBindingRequest.

List of bindings that map each rank to a specific logical NPU and its available NICs in the infrastructure.

List of NIC configuration bindings that specify settings and associated physical bindings for each logical NIC.

List of bindings that associate logical infrastructure elements with physical test resources such as chassis or servers.

name of the logical infrastructure device (e.g. the value dse_infra.GenericHost.name was set to)

0-based device index

name of the logical infrastructure component

0-based component index

These components mark the boundary of a region within the infrastructure.

ref to a NIC defined in the chakra infrastructure

auto-populated, can be overriden by user

flows from this nic may be generated by test ports from any of these associated physical bindings

reference to the logical infrastructure element that the physical resource represents

The platform type for this physical binding (e.g., hardware chassis or software server).

with Keysight Hardware platforms

with NCCL + Keysight Software platforms

Layer 1 physical settings such as link speed, duplex mode, and physical layer parameters.

BEGIN PRIVATE packet capture settings END PRIVATE

The infrastructure region to which a platform will be assigned. Defines the boundary components that comprise the region.

The platform type to assign to this infrastructure region for testing.

ref to NPU used by this rank

The unique identifier for this rank in the distributed workload. Specifies which rank is bound to the referenced NPU.

list of NICs available to NPU (in the same host)