CHANGELOG.md

Changelog for TransferBench

Documentation for TransferBench is available at https://rocm.docs.amd.com/projects/TransferBench.

v1.61.00

Added

• Added a2a_n preset which conducts alltoall GPU-to-GPU tranfers over nearest NIC executors
• Re-implemented GFX_BLOCK_ORDER which allows for control over how threadblocks of multiple transfers are ordered
  • 0 = sequential, 1 = interleaved, 2 = random
• Added a2asweep preset which tries various CU/unroll options for GFX-executed all-to-all
• Rewrite main GID index detection logic
• Show the GID index and description in the topology table. It is helpful for debugging purposes
• Added GFX_WORD_SIZE to allow for different packed float sizes to use for GFX kernel. Must be either 4 (default), 2 or 1

Fixed

• Avoid build errors for CMake and Makefile if infiniband/verbs.h header is not present and disable NIC executor in such case
• Have a priority list of which GID entry to go for instead of hardcoding choices based on underdocumented user input (such as RoCE version and IP address family)
• Use link-local when it is the only choice (i.e. when routing information is not available beyond local link)

v1.60.00

Modified

• Reverted GFX_SINGLE_TEAM default back to 1

Fixed

• Fixed bug where peer memory access was not enabled for DMA transfers, which would break specific DMA engine transfers

v1.59.01

Added

• The a2a preset A2A_MODE variable has been enhanced to allow for customizing the number of srcs/dsts to use This is specified by setting A2A_MODE to numSrcs:numDsts. Extra destinations past 1 will be "local" writes (i.e. if one sets A2A_MODE=1:3, then transfers will follow this pattern: Fx Gx FyFxFx) to simulate similar conditions normally used during collective algorithms such as ring-based AllReduce

v1.59.00

Added

• Adding in support for NIC executor, which allows for RDMA copies on NICs that support IBVerbs By default, NIC executor will be enabled if IBVerbs is found in the dynamic linker cache
• NIC executor can be indexed in two methods
  • "I" Ix.y will use NIC x as the source and NIC y as the destination. E.g. (G0 I0.5 G4)
  • "N" Nx.y will use NIC closest to GPU x as source, and NIC closest to GPU y as destination E.g. (G0 N0.4 N4)
• The closest NIC can be overridden by the environment variable CLOSEST_NIC, which should be a comma-separated list of NIC indices to use for the corresponding GPU
• This feature can be explicitly disabled at compile time by specifying DISABLE_NIC_EXEC=1

Modified

• Changing default data size to 256M from 64M
• Adding NUM_QUEUE_PAIRS which enables NIC traffic in A2A. Each GPU will talk to the next GPU via the closest NIC
• Sweep preset now saves last sweep run configuration to /tmp/lastSweep.cfg and can be changed via SWEEP_FILE

Fixed

• Fixed bug with reporting when using subiterations
• Fixed bug with per-Transfer data size specification
• Fixed bug when using XCC prefered table

v1.58.00

Fixed

• Fixed broken specific DMA-engine copies

v1.57.01

Added

• Re-added "scaling" GPU GFX preset benchmark, which tests copies from GPU to other devices using varying number of CUs.

v1.57.00

Modified

• Removing use of default starship operator / C++20 requirement to enable compilation of more OSs
• Changing how version is reported. Client version is now just last two digits, and increments only if no changes are made to the backend header-only library file, and resets to 0 when header is updated
• GFX_SINGLE_TEAM=0 is set by default

v1.56

Fixed

• Fixed bug when using interactive mode. Interactive mode now starts prior to all warmup iterations

v1.55

Fixed

• Fixed missing header error when compiling on CentOS
• Fixed issues when using multi-stream mode for GFX executor

v1.54

Modified

• Refactored TransferBench into a header-only library combined with a thin client to facilitate the use of TransferBench as the backend for other applications
• Optimized how data validation is handled - this should speed up Tests with many parallel transfers as data is only generated once
• Preset benchmarks now no longer take in any extra command line arguments. Preset settings are only accessed via environment variables. Details for each preset are printed
• The a2a preset benchmark now defaults to using fine-grained memory and GFX unroll of 2
• Refactored how Transfers are launched in parallel which has reduced some CPU-side overheads
• CPU and DMA executor timing now use CPU wall clock timing instead of slowest Transfer time

Added

• New one2all preset which sweeps over all subests of parallel transfers from one GPU to others
• Adding new warnings for DMA execution relating to how HIP will default to using agents from the source memory

Removed

• CU scaling preset has been removed. Similar functionality already exists in the schmoo preset benchmark
• Preparation of source data via GFX kernel has been removed (USE_PREP_KERNEL)
• Removed GFX block-reordering (BLOCK_ORDER)
• Removed NUM_CPU_DEVICES and NUM_GPU_DEVICES from common env vars and only into the presets they apply to.
• Removed SHARED_MEM_BYTES option for GFX executor
• Removed USE_PCIE_INDEX, and SHARED_MEM_BYTES

Fixed

• Fixed a potential timing reporting issue when DMA executed Transfers end up getting serialized.

v1.53

Added

• Added ability to specify NULL for sweep preset as source or destination memory type

v1.52

Added

• Added USE_HSA_DMA env var to switch to using hsa_amd_memory_async_copy instead of hipMemcpyAsync for DMA execution
• Added ability to set USE_GPU_DMA env var for a2a benchmark
• Adding check for large BAR enablement for GPU devices during topology check

Fixed

• Potential memory leak if HSA reports 0 hops between GPUs and CPUs

v1.51

Modified

• CSV output has been modified slightly to match normal terminal output
• Output for non single stream mode has been changed to match single stream mode (results per Executor)

Added

• Support for sub-iterations via NUM_SUBITERATIONS. This allows for additional looping during an iteration If set to 0, this should infinitely loop (which may be useful for some debug purposes)
• Support for variable number of subexecutors (currently for GPU-GFX executor only). Setting subExecutors to 0 will run over a range of CUs to use, and report only the results of the best one found. This can be tuned for performance by setting the MIN_VAR_SUBEXEC and MAX_VAR_SUBEXEC environment variables to narrow the search space. The number of CUs used will be identical for all variable subExecutor transfers
• Experimental new "healthcheck" preset config which currently only supports MI300 series. This preset runs through CPU to GPU bandwidth tests and all-to-all XGMI bandwidth tests and compares against expected values Pass criteria limits can be modified (due to platform differences) via the environment variables LIMIT_UDIR (undirectional), LIMIT_BDIR (bidirectional), and LIMIT_A2A (Per GPU-GPU link bandwidth)

Fixed

• Fixed out-of-bounds memory access during topology detection that can happen if the number of CPUs is less than the number of NUMA domains
• Fixed CU masking functionality on multi-XCD architectures (e.g. MI300)

v1.50

Added

• Adding new parallel copy preset benchmark (pcopy)
  • Usage: ./TransferBench pcopy <numBytes=64M> <#CUs=8> <srcGpu=0> <minGpus=1> <maxGpus=#GPU-1>

Fixed

• Removed non-copies DMA Transfers (this had previously been using hipMemset)
• Fixed CPU executor when operating on null destination

v1.49

Fixes

• Enumerating previously missed DMA engines used only for CPU traffic in topology display

v1.48

Fixes

• Various fixes for TransferBenchCuda

Additions

• Support for targeting specific DMA engines via executor subindex (e.g. D0.1)
• Printing warnings when exeuctors are overcommited

Modifications

• USE_REMOTE_READ supported for rwrite preset benchmark

v1.47

Fixes

• Fixing CUDA support

v1.46

Fixes

• Fixing GFX_UNROLL set to 13 (past 8) on gfx906 cards

Modifications

• GFX_SINGLE_TEAM=1 by default
• Adding field showing summation of individual Transfer bandwidths for Executors

v1.45

Additions

• Adding A2A_MODE to a2a preset (0 = copy, 1 = read-only, 2 = write-only)
• Adding GFX_UNROLL to modify GFX kernel's unroll factor
• Adding GFX_WAVE_ORDER to modify order in which wavefronts process data

Modifications

• Rewrote the GFX reduction kernel to support new wave ordering

v1.44

Additions

• Adding rwrite preset to benchmark remote parallel writes
• Usage: ./TransferBench rwrite <numBytes=64M> <#CUs=8> <srcGpu=0> <minGpus=1> <maxGpus=3>

v1.43

Changes

• Modifying a2a to show executor timing, as well as executor min/max bandwidth

v1.42

Fixes

• Fixing schmoo maxNumCus optional arg parsing
• Schmoo output modified to be easier to copy

v1.41

Additions

• Adding schmoo preset config benchmarks local/remote reads/writes/copies
  • Usage: ./TransferBench schmoo <numBytes=64M> <localIdx=0> <remoteIdx=1> <maxNumCUs=32>

Fixes

• Fixing some misreported timings when running with non-fixed number of iterations

v1.40

Fixes

• Fixing XCC defaulting to 0 instead of random for preset configs, ignoring XCC_PREF_TABLE

v1.39

Additions

• (Experimental) Adding support for Executor sub-index

Fixes

• Remove deprecated gcnArch code. ROCm version must include support for hipDeviceMallocUncached

v1.38

Fixes

• Adding missing threadfence which could cause non-fine-grained Transfers to report higher speeds

v1.37

Changes

• USE_SINGLE_STREAM is enabled by default now. (Disable via USE_SINGLE_STREAM=0)

Fixes

• Fix unrecognized token error when XCC_PREF_TABLE is unspecified

v1.36

Additions

• (Experimental) Adding XCC filtering - combined with XCC_PREF_TABLE, this tries to select specific XCCs to use for specific (SRC->DST) Transfers

v1.35

Additions

• USE_FINE_GRAIN also applies to a2a preset

v1.34

Additions

• Set GPU_KERNEL=3 as default for gfx942

v1.33

Additions

• Added the ALWAYS_VALIDATE environment variable to allow for validation after every iteration, instead of only once at the end of all iterations

v1.32

Changes

• Increased the line limit from 2048 to 32768

v1.31

Changes

• SHOW_ITERATIONS now shows XCC:CU instead of just CU ID
• SHOW_ITERATIONS is printed when USE_SINGLE_STREAM=1

v1.30

Additions

• BLOCK_SIZE has been added to control the threadblock size (must be a multiple of 64, up to 512)
• BLOCK_ORDER has been added to control how work is ordered for GFX-executors running USE_SINGLE_STREAM=1
  • 0 - Threadblocks for transfers are ordered sequentially (default)
  • 1 - Threadblocks for transfers are interleaved
  • 2 - Threadblocks for transfers are ordered randomly

v1.29

Additions

• A2A preset config now responds to USE_REMOTE_READ

Fixes

• Race-condition during wall-clock initialization caused "inf" during single-stream runs
• CU numbering output after CU masking

Changes

• The default number of warmups has been reverted to 3
• The default unroll factor for gfx940/941 has been set to 6

v1.28

Additions

• Added A2A_DIRECT, which only runs all-to-all on directly connected GPUs (now on by default)
• Added average statistics for P2P and A2A benchmarks
• Added USE_FINE_GRAIN for P2P benchmark
  • With older devices, P2P performance with default coarse-grain device memory stops timing as soon as a request is sent to data fabric, and not actually when it arrives remotely. This can artificially inflate bandwidth numbers, especially when sending small amounts of data.

Changes

• Modified P2P output to help distinguish between CPU and GPU devices

Fixes

• Fixed Makefile target to prevent unnecessary re-compilation

v1.27

Additions

• Added cmdline preset to allow specification of simple tests on command line (e.g., ./TransferBench cmdline 64M "1 4 G0->G0->G1")
• Adding the HIDE_ENV environment variable, which stops environment variable values from printing
• Adding the CU_MASK environment variable, which allows you to select the CUs to run on
• CU_MASK is specified in CU indices (0-#CUs-1), where ' - ' can be used to denote ranges of values (e.g., CU_MASK=3-8,16 requests that transfer be run only on CUs 3,4,5,6,7,8,16)
  • Note that this is somewhat experimental and may not work on all hardware
• SHOW_ITERATIONS now shows CU usage for that iteration (experimental)

Changes

• Added extra comments on commonly missing includes with details on how to install them

Fixes

• CUDA compilation works again (the wall_clock64 CUDA alias was not defined)

v1.26

Additions

• Setting SHOW_ITERATIONS=1 provides additional information about per-iteration timing for file and P2P configs
  • For file configs, iterations are sorted from min to max bandwidth and displayed with standard deviation
  • For P2P, min/max/standard deviation is shown for each direction

Changes

• P2P benchmark formatting now reports bidirectional bandwidth in each direction (as well as sum) for clarity

v1.25

Fixes

• Fixed a bug in the P2P bidirectional benchmark that used the incorrect number of subExecutors for CPU<->GPU tests

v1.24

Additions

• New All-To-All GPU benchmark accessed by preset "A2A"
• Added gfx941 wall clock frequency

v1.23

Additions

• New GPU subexec scaling benchmark accessed by preset "scaling"
  • Tests GPU-GFX copy performance based on # of CUs used

v1.22

Changes

• Switched the kernel timing function to wall_clock64

v1.21

Fixes

• Fixed a bug with SAMPLING_FACTOR

v1.20

Fixes

• VALIDATE_DIRECT can now be used with USE_PREP_KERNEL
• Switched to local GPU for validating GPU memory

v1.19

Additions

• VALIDATE_DIRECT now also applies to source memory array checking
• Added null memory pointer check prior to deallocation

v1.18

Additions

• Adding the ability to validate GPU destination memory directly without going through the CPU staging buffer (VALIDATE_DIRECT)
  • Note that this only works on AMD devices with large-bar access enabled, and may slow things down considerably

Changes

• Refactored how environment variables are displayed
• Mismatch stops after the first detected error within an array instead of listing all mismatched elements

v1.17

Additions

• Allowed switch to GFX kernel for source array initialization (USE_PREP_KERNEL)
  • Note that USE_PREP_KERNEL can't be used with FILL_PATTERN
• Added the ability to compile with nvcc only (TransferBenchCuda)

Changes

• The default pattern was set to [Element i = ((i * 517) modulo 383 + 31) * (srcBufferIdx + 1)]

Fixes

• Added the example.cfg file

v1.16

Additions

• Additional src array validation during preparation
• Added a new environment variable (CONTINUE_ON_ERROR) to resume tests after a mis-match detection
• Initialized GPU memory to 0 during allocation

v1.15

Fixes

• Fixed a bug that prevented single transfers greater than 8 GB

Changes

• Removed "check for latest ROCm" warning when allocating too much memory
• Off-source memory value is now printed when a mis-match is detected

v1.14

Additions

• Added documentation
• Added pthread linking in src/Makefile and CMakeLists.txt
• Added printing off the hex value of the floats for output and reference

v1.13

Additions

• Added support for cmake

Changes

• Converted to the Pitchfork layout standard

v1.12

Additions

• Added support for TransferBench on NVIDIA platforms (via HIP_PLATFORM=nvidia)
  • Note that CPU executors on NVIDIA platform cannot access GPU memory (no large-bar access)

v1.11

Additions

• Added multi-input/multi-output (MIMO) support: transfers now can reduce (element-wise summation) multiple input memory arrays and write sums to multiple outputs
• Added GPU-DMA executor 'D', which uses hipMemcpy for SDMA copies
  • Previously, this was done using USE_HIP_CALL, but now GPU-GFX kernel can run in parallel with GPU-DMA, instead of applying to all GPU executors globally
  • GPU-DMA executor can only be used for single-input/single-output transfers
  • GPU-DMA executor can only be associated with one SubExecutor
• Added new "Null" memory type 'N', which represents empty memory. This allows for read-only or write-only transfers
• Added new GPU_KERNEL environment variable that allows switching between various GPU-GFX reduction kernels

Optimizations

• Improved GPU-GFX kernel performance based on hardware architecture when running with fewer CUs

Changes

• Updated the example.cfg file to cover new features
• Updated output to support MIMO
• Changed CU and CPU thread naming to SubExecutors for consistency
• Sweep Preset: default sweep preset executors now includes DMA
• P2P benchmarks:
  • Removed p2p_rr, g2g and g2g_rr (now only works via P2P)
    • Setting NUM_CPU_DEVICES=0 can only be used to benchmark GPU devices (like g2g)
    • The new USE_REMOTE_READ environment variable replaces _rr presets
  • New environment variable USE_GPU_DMA=1 replaces USE_HIP_CALL=1 for benchmarking with GPU-DMA Executor
  • Number of GPU SubExecutors for benchmark can be specified via NUM_GPU_SE
    • Defaults to all CUs for GPU-GFX, 1 for GPU-DMA
  • Number of CPU SubExecutors for benchmark can be specified via NUM_CPU_SE
• Psuedo-random input pattern has been slightly adjusted to have different patterns for each input array within same transfer

Removals

• USE_HIP_CALL: use GPU-DMA executor 'D' or set USE_GPU_DMA=1 for P2P benchmark presets
  • Currently, a warning will be issued if USE_HIP_CALL is set to 1 and the program will stop
• NUM_CPU_PER_TRANSFER: the number of CPU SubExecutors will be whatever is specified for the transfer
• USE_MEMSET: this function can now be done via a transfer using the null memory type

v1.10

Fixes

• Fixed incorrect bandwidth calculation when using single stream mode and per-transfer data sizes

v1.09

Additions

• Printing off src/dst memory addresses during interactive mode

Changes

• Switching to numa_set_preferred instead of set_mempolicy

v1.08

Changes

• Fixed handling of non-configured NUMA nodes
• Topology detection now shows actual NUMA node indices
• Fixed 'for' issue with NUM_GPU_DEVICES

v1.07

Fixes

• Fixed bug with allocations involving non-default CPU memory types

v1.06

Additions

• Unpinned CPU memory type ('U'), which may require HSA_XNACK=1 in order to access via GPU executors
• Added sweep configuration logging to lastSweep.cfg
• Ability to specify the number of CUs to use for sweep-based presets

Changes

• Modified advanced configuration file format to accept bytes-per-transfer

Fixes

• Fixed random sweep repeatability
• Fixed bug with CPU NUMA node memory allocation

v1.05

Additions

• Topology output now includes NUMA node information
• Support for NUMA nodes with no CPU cores (e.g., CXL memory)

Removals

• The SWEEP_SRC_IS_EXE environment variable was removed

v1.04

Additions

• There are new environment variables for sweep based presets:
  • SWEEP_XGMI_MIN: The minumum number of XGMI hops for transfers
  • SWEEP_XGMI_MAX: The maximum number of XGMI hops for transfers
  • SWEEP_SEED: Uses a random seed
  • SWEEP_RAND_BYTES: Uses a random amount of bytes (up to pre-specified N) for each transfer

Changes

• CSV output for sweep now includes an environment variables section followed by output
• CSV output no longer lists environment variable parameters in columns
• We changed the default number of warmup iterations from 3 to 1
• Split CSV output of link type to ExeToSrcLinkType and ExeToDstLinkType

v1.03

Additions

• There are new preset modes stress-test benchmarks: sweep and randomsweep
  • sweep iterates over all possible sets of transfers to test
  • randomsweep iterates over random sets of transfers
  • New sweep-only environment variables can modify sweep
    • SWEEP_SRC: String containing only "B","C","F", or "G" that defines possible source memory types
    • SWEEP_EXE: String containing only "C" or "G" that defines possible executors
    • SWEEP_DST: String containing only "B","C","F", or "G" that defines possible destination memory types
    • SWEEP_SRC_IS_EXE: Restrict the executor to be the same as the source, if non-zero
    • SWEEP_MIN: Minimum number of parallel transfers to test
    • SWEEP_MAX: Maximum number of parallel transfers to test
    • SWEEP_COUNT: Maximum number of tests to run
    • SWEEP_TIME_LIMIT: Maximum number of seconds to run tests
• New environment variables to restrict number of available devices to test on (primarily for sweep runs)
  • NUM_CPU_DEVICES: Number of CPU devices
  • NUM_GPU_DEVICES: Number of GPU devices

Fixes

• Fixed timing display for CPU executors when using single-stream mode

v1.02

Additions

• Setting NUM_ITERATIONS to a negative number indicates a run of -NUM_ITERATIONS seconds per test

Changes

• Copies are now referred to as 'transfers' instead of 'links'
• Reordered how environment variables are displayed (alphabetically now)

Removals

• Combined timing is now always on for kernel-based GPU copies; the COMBINED_TIMING environment variable has been removed
• Single sync is no longer supported for facility variable iterations; the USE_SINGLE_SYNC environmental variable has been removed

v1.01

Additions

• Added the USE_SINGLE_STREAM feature
  • All Links that run on the same GPU device are run with a single kernel launch on a single stream
  • This doesn't work with USE_HIP_CALL, and it forces USE_SINGLE_SYNC to collect timings
  • Added the ability to request coherent or fine-grained host memory ('B')

Changes

• Separated the TransferBench repository from the RCCL repository
• Peer-to-peer benchmark mode now works with OUTPUT_TO_CSV
• Toplogy display now works with OUTPUT_TO_CSV
• Moved the documentation about the config file into example.cfg

Removals

• Removed config file generation
• Removed the 'show pointer address' (SHOW_ADDR) environment variable