Skip to content

GitLab

Unverified Commit cd80b3a3 authored by gilbertlee-amd's avatar gilbertlee-amd Committed by GitHub
Browse files

TransferBench v1.61 (#174) 


Co-authored-by: default avatarMustafa Abduljabbar <mustafa.abduljabbar@amd.com>
parent 856e3445
Expand all Show whitespace changes
Inline Side-by-side
Showing with 821 additions and 309 deletions
CHANGELOG.md
View file @ cd80b3a3
......@@ -3,6 +3,22 @@
Documentation for TransferBench is available at
[https://rocm.docs.amd.com/projects/TransferBench](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
......
CMakeLists.txt
View file @ cd80b3a3
......@@ -57,17 +57,22 @@ set( CMAKE_CXX_FLAGS "${flags_str} ${CMAKE_CXX_FLAGS}")
set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3 -L${ROCM_PATH}/lib")
include_directories(${ROCM_PATH}/include)
find_library(IBVERBS_LIBRARY ibverbs)
if (IBVERBS_LIBRARY)
if (DEFINED ENV{DISABLE_NIC_EXEC})
find_path(IBVERBS_INCLUDE_DIR infiniband/verbs.h)
if (DEFINED ENV{DISABLE_NIC_EXEC})
message(STATUS "Disabling NIC Executor support")
else()
elseif(IBVERBS_LIBRARY AND IBVERBS_INCLUDE_DIR)
message(STATUS "Found ibverbs: ${IBVERBS_LIBRARY}. Building with NIC executor support. Can set DISABLE_NIC_EXEC=1 to disable")
add_definitions(-DNIC_EXEC_ENABLED)
link_libraries(ibverbs)
endif()
else()
message(WARNING "IBVerbs library not found. Building without NIC executor support")
if (NOT IBVERBS_LIBRARY)
message(WARNING "IBVerbs library not found")
elseif (NOT IBVERBS_INCLUDE_DIR)
message(WARNING "infiniband/verbs.h not found")
endif()
message(WARNING "Building without NIC executor support. To use the TransferBench RDMA executor, check if your system has NICs, the NIC drivers are installed, and libibverbs-dev is installed")
endif()
link_libraries(numa hsa-runtime64 pthread)
set (CMAKE_RUNTIME_OUTPUT_DIRECTORY .)
add_executable(TransferBench src/client/Client.cpp)
......
Makefile
View file @ cd80b3a3
#
# Copyright (c) 2019-2024 Advanced Micro Devices, Inc. All rights reserved.
# Copyright (c) 2019-2025 Advanced Micro Devices, Inc. All rights reserved.
#
# Configuration options
......@@ -12,8 +12,10 @@ NVCC=$(CUDA_PATH)/bin/nvcc
# Compile TransferBenchCuda if nvcc detected
ifeq ("$(shell test -e $(NVCC) && echo found)", "found")
EXE=TransferBenchCuda
CXX=$(NVCC)
else
EXE=TransferBench
CXX=$(HIPCC)
endif
CXXFLAGS = -I$(ROCM_PATH)/include -lnuma -L$(ROCM_PATH)/lib -lhsa-runtime64
......@@ -21,14 +23,24 @@ NVFLAGS = -x cu -lnuma -arch=native
COMMON_FLAGS = -O3 -I./src/header -I./src/client -I./src/client/Presets
LDFLAGS += -lpthread
# Compile RDMA executor if IBVerbs is found in the Dynamic Linker cache
# Compile RDMA executor if
# 1) DISABLE_NIC_EXEC is not set to 1
# 2) IBVerbs is found in the Dynamic Linker cache
# 3) infiniband/verbs.h is found in the default include path
NIC_ENABLED = 0
ifneq ($(DISABLE_NIC_EXEC),1)
ifneq ("$(shell ldconfig -p | grep -c ibverbs)", "0")
ifeq ("$(shell ldconfig -p | grep -c ibverbs)", "0")
$(info lib IBVerbs not found)
else ifeq ("$(shell echo '#include <infiniband/verbs.h>' | $(CXX) -E - 2>/dev/null | grep -c 'infiniband/verbs.h')", "0")
$(info infiniband/verbs.h not found)
else
LDFLAGS += -libverbs -DNIC_EXEC_ENABLED
NVFLAGS += -libverbs -DNIC_EXEC_ENABLED
NIC_ENABLED = 1
endif
ifeq ($(NIC_ENABLED), 0)
$(info To use the TransferBench RDMA executor, check if your system has NICs, the NIC drivers are installed, and libibverbs-dev is installed)
endif
endif
all: $(EXE)
......
src/client/EnvVars.hpp
View file @ cd80b3a3
/*
Copyright (c) 2021-2024 Advanced Micro Devices, Inc. All rights reserved.
Copyright (c) 2021-2025 Advanced Micro Devices, Inc. All rights reserved.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
......@@ -84,6 +84,7 @@ public:
int useHsaDma; // Use hsa_amd_async_copy instead of hipMemcpy for non-targetted DMA executions
// GFX options
int gfxBlockOrder; // How threadblocks for multiple Transfers are ordered 0=sequential 1=interleaved
int gfxBlockSize; // Size of each threadblock (must be multiple of 64)
vector<uint32_t> cuMask; // Bit-vector representing the CU mask
vector<vector<int>> prefXccTable; // Specifies XCC to use for given exe->dst pair
......@@ -92,6 +93,7 @@ public:
int useSingleStream; // Use a single stream per GPU GFX executor instead of stream per Transfer
int gfxSingleTeam; // Team all subExecutors across the data array
int gfxWaveOrder; // GFX-kernel wavefront ordering
int gfxWordSize; // GFX-kernel packed data size (4=DWORDx4, 2=DWORDx2, 1=DWORDx1)
// Client options
int hideEnv; // Skip printing environment variable
......@@ -135,10 +137,12 @@ public:
alwaysValidate = GetEnvVar("ALWAYS_VALIDATE" , 0);
blockBytes = GetEnvVar("BLOCK_BYTES" , 256);
byteOffset = GetEnvVar("BYTE_OFFSET" , 0);
gfxBlockOrder = GetEnvVar("GFX_BLOCK_ORDER" , 0);
gfxBlockSize = GetEnvVar("GFX_BLOCK_SIZE" , 256);
gfxSingleTeam = GetEnvVar("GFX_SINGLE_TEAM" , 1);
gfxUnroll = GetEnvVar("GFX_UNROLL" , defaultGfxUnroll);
gfxWaveOrder = GetEnvVar("GFX_WAVE_ORDER" , 0);
gfxWordSize = GetEnvVar("GFX_WORD_SIZE" , 4);
hideEnv = GetEnvVar("HIDE_ENV" , 0);
minNumVarSubExec = GetEnvVar("MIN_VAR_SUBEXEC" , 1);
maxNumVarSubExec = GetEnvVar("MAX_VAR_SUBEXEC" , 0);
......@@ -286,13 +290,23 @@ public:
}
}
static std::string ToStr(std::vector<int> const& values) {
std::string result = "";
bool isFirst = true;
for (int v : values) {
if (isFirst) isFirst = false;
else result += ",";
result += std::to_string(v);
}
return result;
}
// Display info on the env vars that can be used
static void DisplayUsage()
{
printf("Environment variables:\n");
printf("======================\n");
printf(" ALWAYS_VALIDATE - Validate after each iteration instead of once after all iterations\n");
printf(" BLOCK_SIZE - # of threads per threadblock (Must be multiple of 64)\n");
printf(" BLOCK_BYTES - Controls granularity of how work is divided across subExecutors\n");
printf(" BYTE_OFFSET - Initial byte-offset for memory allocations. Must be multiple of 4\n");
#if NIC_EXEC_ENABLED
......@@ -300,9 +314,12 @@ public:
#endif
printf(" CU_MASK - CU mask for streams. Can specify ranges e.g '5,10-12,14'\n");
printf(" FILL_PATTERN - Big-endian pattern for source data, specified in hex digits. Must be even # of digits\n");
printf(" GFX_BLOCK_ORDER - How blocks for transfers are ordered. 0=sequential, 1=interleaved\n");
printf(" GFX_BLOCK_SIZE - # of threads per threadblock (Must be multiple of 64)\n");
printf(" GFX_UNROLL - Unroll factor for GFX kernel (0=auto), must be less than %d\n", TransferBench::GetIntAttribute(ATR_GFX_MAX_UNROLL));
printf(" GFX_SINGLE_TEAM - Have subexecutors work together on full array instead of working on disjoint subarrays\n");
printf(" GFX_WAVE_ORDER - Stride pattern for GFX kernel (0=UWC,1=UCW,2=WUC,3=WCU,4=CUW,5=CWU)\n");
printf(" GFX_WORD_SIZE - GFX kernel packed data size (4=DWORDx4, 2=DWORDx2, 1=DWORDx1)\n");
printf(" HIDE_ENV - Hide environment variable value listing\n");
#if NIC_EXEC_ENABLED
printf(" IB_GID_INDEX - Required for RoCE NICs (default=-1/auto)\n");
......@@ -383,6 +400,8 @@ public:
"%s", (cuMask.size() ? GetCuMaskDesc().c_str() : "All"));
Print("FILL_PATTERN", getenv("FILL_PATTERN") ? 1 : 0,
"%s", (fillPattern.size() ? getenv("FILL_PATTERN") : TransferBench::GetStrAttribute(ATR_SRC_PREP_DESCRIPTION).c_str()));
Print("GFX_BLOCK_ORDER", gfxBlockOrder,
"Thread block ordering: %s", gfxBlockOrder == 0 ? "Sequential" : "Interleaved");
Print("GFX_BLOCK_SIZE", gfxBlockSize,
"Threadblock size of %d", gfxBlockSize);
Print("GFX_SINGLE_TEAM", gfxSingleTeam,
......@@ -397,6 +416,9 @@ public:
gfxWaveOrder == 3 ? "Wavefront,CU,Unroll" :
gfxWaveOrder == 4 ? "CU,Unroll,Wavefront" :
"CU,Wavefront,Unroll"));
Print("GFX_WORD_SIZE", gfxWordSize,
"Using GFX word size of %d (DWORDx%d)", gfxWordSize, gfxWordSize);
#if NIC_EXEC_ENABLED
Print("IP_ADDRESS_FAMILY", ipAddressFamily,
"IP address family is set to IPv%d", ipAddressFamily);
......@@ -462,6 +484,31 @@ public:
return defaultValue;
}
static std::vector<int> GetEnvVarArray(std::string const& varname, std::vector<int> const& defaultValue)
{
if (getenv(varname.c_str())) {
char* rangeStr = getenv(varname.c_str());
std::set<int> values;
char* token = strtok(rangeStr, ",");
while (token) {
int start, end;
if (sscanf(token, "%d-%d", &start, &end) == 2) {
for (int i = start; i <= end; i++) values.insert(i);
} else if (sscanf(token, "%d", &start) == 1) {
values.insert(start);
} else {
printf("[ERROR] Unrecognized token [%s]\n", token);
exit(1);
}
token = strtok(NULL, ",");
}
std::vector<int> result;
for (auto v : values) result.push_back(v);
return result;
}
return defaultValue;
}
static std::string GetEnvVar(std::string const& varname, std::string const& defaultValue)
{
if (getenv(varname.c_str()))
......@@ -524,6 +571,7 @@ public:
cfg.dma.useHipEvents = useHipEvents;
cfg.dma.useHsaCopy = useHsaDma;
cfg.gfx.blockOrder = gfxBlockOrder;
cfg.gfx.blockSize = gfxBlockSize;
cfg.gfx.cuMask = cuMask;
cfg.gfx.prefXccTable = prefXccTable;
......@@ -532,6 +580,7 @@ public:
cfg.gfx.useMultiStream = !useSingleStream;
cfg.gfx.useSingleTeam = gfxSingleTeam;
cfg.gfx.waveOrder = gfxWaveOrder;
cfg.gfx.wordSize = gfxWordSize;
cfg.nic.ibGidIndex = ibGidIndex;
cfg.nic.ibPort = ibPort;
......
src/client/Presets/AllToAllN.hpp 0 → 100644
View file @ cd80b3a3
/*
Copyright (c) 2024 Advanced Micro Devices, Inc. All rights reserved.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
*/
#include "EnvVars.hpp"
void AllToAllRdmaPreset(EnvVars& ev,
size_t const numBytesPerTransfer,
std::string const presetName)
{
int numDetectedGpus = TransferBench::GetNumExecutors(EXE_GPU_GFX);
// Collect env vars for this preset
int numGpus = EnvVars::GetEnvVar("NUM_GPU_DEVICES", numDetectedGpus);
int numQueuePairs = EnvVars::GetEnvVar("NUM_QUEUE_PAIRS", 1);
int useFineGrain = EnvVars::GetEnvVar("USE_FINE_GRAIN" , 1);
// Print off environment variables
ev.DisplayEnvVars();
if (!ev.hideEnv) {
if (!ev.outputToCsv) printf("[AllToAll Network Related]\n");
ev.Print("NUM_GPU_DEVICES", numGpus , "Using %d GPUs", numGpus);
ev.Print("NUM_QUEUE_PAIRS", numQueuePairs, "Using %d queue pairs for NIC transfers", numQueuePairs);
ev.Print("USE_FINE_GRAIN" , useFineGrain , "Using %s-grained memory", useFineGrain ? "fine" : "coarse");
printf("\n");
}
// Validate env vars
if (numGpus < 0 || numGpus > numDetectedGpus) {
printf("[ERROR] Cannot use %d GPUs. Detected %d GPUs\n", numGpus, numDetectedGpus);
exit(1);
}
MemType memType = useFineGrain ? MEM_GPU_FINE : MEM_GPU;
std::map<std::pair<int, int>, int> reIndex;
std::vector<Transfer> transfers;
for (int i = 0; i < numGpus; i++) {
for (int j = 0; j < numGpus; j++) {
// Build Transfer and add it to list
TransferBench::Transfer transfer;
transfer.numBytes = numBytesPerTransfer;
transfer.srcs.push_back({memType, i});
transfer.dsts.push_back({memType, j});
transfer.exeDevice = {EXE_NIC_NEAREST, i};
transfer.exeSubIndex = j;
transfer.numSubExecs = numQueuePairs;
reIndex[std::make_pair(i,j)] = transfers.size();
transfers.push_back(transfer);
}
}
printf("GPU-RDMA All-To-All benchmark:\n");
printf("==========================\n");
printf("- Copying %lu bytes between all pairs of GPUs using %d QPs per Transfer (%lu Transfers)\n",
numBytesPerTransfer, numQueuePairs, transfers.size());
if (transfers.size() == 0) return;
// Execute Transfers
TransferBench::ConfigOptions cfg = ev.ToConfigOptions();
TransferBench::TestResults results;
if (!TransferBench::RunTransfers(cfg, transfers, results)) {
for (auto const& err : results.errResults)
printf("%s\n", err.errMsg.c_str());
exit(0);
} else {
PrintResults(ev, 1, transfers, results);
}
// Print results
char separator = (ev.outputToCsv ? ',' : ' ');
printf("\nSummary: [%lu bytes per Transfer]\n", numBytesPerTransfer);
printf("==========================================================\n");
printf("SRC\\DST ");
for (int dst = 0; dst < numGpus; dst++)
printf("%cGPU %02d ", separator, dst);
printf(" %cSTotal %cActual\n", separator, separator);
double totalBandwidthGpu = 0.0;
double minActualBandwidth = std::numeric_limits<double>::max();
double maxActualBandwidth = 0.0;
std::vector<double> colTotalBandwidth(numGpus+2, 0.0);
for (int src = 0; src < numGpus; src++) {
double rowTotalBandwidth = 0;
int transferCount = 0;
double minBandwidth = std::numeric_limits<double>::max();
printf("GPU %02d", src);
for (int dst = 0; dst < numGpus; dst++) {
if (reIndex.count(std::make_pair(src, dst))) {
int const transferIdx = reIndex[std::make_pair(src,dst)];
TransferBench::TransferResult const& r = results.tfrResults[transferIdx];
colTotalBandwidth[dst] += r.avgBandwidthGbPerSec;
rowTotalBandwidth += r.avgBandwidthGbPerSec;
totalBandwidthGpu += r.avgBandwidthGbPerSec;
minBandwidth = std::min(minBandwidth, r.avgBandwidthGbPerSec);
transferCount++;
printf("%c%8.3f ", separator, r.avgBandwidthGbPerSec);
} else {
printf("%c%8s ", separator, "N/A");
}
}
double actualBandwidth = minBandwidth * transferCount;
printf(" %c%8.3f %c%8.3f\n", separator, rowTotalBandwidth, separator, actualBandwidth);
minActualBandwidth = std::min(minActualBandwidth, actualBandwidth);
maxActualBandwidth = std::max(maxActualBandwidth, actualBandwidth);
colTotalBandwidth[numGpus+1] += rowTotalBandwidth;
}
printf("\nRTotal");
for (int dst = 0; dst < numGpus; dst++) {
printf("%c%8.3f ", separator, colTotalBandwidth[dst]);
}
printf(" %c%8.3f %c%8.3f %c%8.3f\n", separator, colTotalBandwidth[numGpus+1],
separator, minActualBandwidth, separator, maxActualBandwidth);
printf("\n");
printf("Average bandwidth (Tx Thread Timed): %8.3f GB/s\n", totalBandwidthGpu / transfers.size());
printf("Aggregate bandwidth (Tx Thread Timed): %8.3f GB/s\n", totalBandwidthGpu);
printf("Aggregate bandwidth (CPU Timed): %8.3f GB/s\n", results.avgTotalBandwidthGbPerSec);
PrintErrors(results.errResults);
}
src/client/Presets/AllToAllSweep.hpp 0 → 100644
View file @ cd80b3a3
/*
Copyright (c) 2025 Advanced Micro Devices, Inc. All rights reserved.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
*/
#include "EnvVars.hpp"
void AllToAllSweepPreset(EnvVars& ev,
size_t const numBytesPerTransfer,
std::string const presetName)
{
enum
{
A2A_COPY = 0,
A2A_READ_ONLY = 1,
A2A_WRITE_ONLY = 2,
A2A_CUSTOM = 3,
};
char a2aModeStr[4][20] = {"Copy", "Read-Only", "Write-Only", "Custom"};
// Force single-stream mode for all-to-all benchmark
ev.useSingleStream = 1;
int numDetectedGpus = TransferBench::GetNumExecutors(EXE_GPU_GFX);
// Collect env vars for this preset
int a2aDirect = EnvVars::GetEnvVar("A2A_DIRECT" , 1);
int a2aLocal = EnvVars::GetEnvVar("A2A_LOCAL" , 0);
int numGpus = EnvVars::GetEnvVar("NUM_GPU_DEVICES", numDetectedGpus);
int useFineGrain = EnvVars::GetEnvVar("USE_FINE_GRAIN" , 1);
int useRemoteRead = EnvVars::GetEnvVar("USE_REMOTE_READ", 0);
int useSpray = EnvVars::GetEnvVar("USE_SPRAY", 0);
int verbose = EnvVars::GetEnvVar("VERBOSE", 0);
std::vector<int> unrollList = EnvVars::GetEnvVarArray("UNROLLS", {1,2,3,4,6,8});
std::vector<int> numCusList = EnvVars::GetEnvVarArray("NUM_CUS", {4,8,12,16,24,32});
// A2A_MODE may be 0,1,2 or else custom numSrcs:numDsts
int numSrcs, numDsts;
int a2aMode = 0;
if (getenv("A2A_MODE") && sscanf(getenv("A2A_MODE"), "%d:%d", &numSrcs, &numDsts) == 2) {
a2aMode = A2A_CUSTOM;
} else {
a2aMode = EnvVars::GetEnvVar("A2A_MODE", 0);
if (a2aMode < 0 || a2aMode > 2) {
printf("[ERROR] a2aMode must be between 0 and 2, or else numSrcs:numDsts\n");
exit(1);
}
numSrcs = (a2aMode == A2A_WRITE_ONLY ? 0 : 1);
numDsts = (a2aMode == A2A_READ_ONLY ? 0 : 1);
}
// Print off environment variables
ev.DisplayEnvVars();
if (!ev.hideEnv) {
if (!ev.outputToCsv) printf("[AllToAll Related]\n");
ev.Print("A2A_DIRECT" , a2aDirect , a2aDirect ? "Only using direct links" : "Full all-to-all");
ev.Print("A2A_LOCAL" , a2aLocal , "%s local transfers", a2aLocal ? "Include" : "Exclude");
ev.Print("A2A_MODE" , (a2aMode == A2A_CUSTOM) ? std::to_string(numSrcs) + ":" + std::to_string(numDsts) : std::to_string(a2aMode),
(a2aMode == A2A_CUSTOM) ? (std::to_string(numSrcs) + " read(s) " +
std::to_string(numDsts) + " write(s)").c_str(): a2aModeStr[a2aMode]);
ev.Print("NUM_CUS" , numCusList.size(), EnvVars::ToStr(numCusList).c_str());
ev.Print("NUM_GPU_DEVICES", numGpus , "Using %d GPUs", numGpus);
ev.Print("UNROLLS" , unrollList.size(), EnvVars::ToStr(unrollList).c_str());
ev.Print("USE_FINE_GRAIN" , useFineGrain , "Using %s-grained memory", useFineGrain ? "fine" : "coarse");
ev.Print("USE_REMOTE_READ", useRemoteRead , "Using %s as executor", useRemoteRead ? "DST" : "SRC");
ev.Print("USE_SPRAY" , useSpray , "%s per CU", useSpray ? "All targets" : "One target");
ev.Print("VERBOSE" , verbose , verbose ? "Display test results" : "Display summary only");
printf("\n");
}
// Validate env vars
if (numGpus < 0 || numGpus > numDetectedGpus) {
printf("[ERROR] Cannot use %d GPUs. Detected %d GPUs\n", numGpus, numDetectedGpus);
exit(1);
}
if (useSpray && numDsts > 1) {
printf("[ERROR] Cannot use USE_SPRAY with multiple destination buffers\n");
exit(1);
}
// Collect the number of GPU devices to use
MemType memType = useFineGrain ? MEM_GPU_FINE : MEM_GPU;
ExeType exeType = EXE_GPU_GFX;
std::vector<Transfer> transfers;
int targetCount = 0;
if (!useSpray) {
// Each CU will work on just one target
for (int i = 0; i < numGpus; i++) {
targetCount = 0;
for (int j = 0; j < numGpus; j++) {
// Check whether or not to execute this pair
if (i == j) {
if (!a2aLocal) continue;
} else if (a2aDirect) {
#if !defined(__NVCC__)
uint32_t linkType, hopCount;
HIP_CALL(hipExtGetLinkTypeAndHopCount(i, j, &linkType, &hopCount));
if (hopCount != 1) continue;
#endif
}
// Build Transfer and add it to list
TransferBench::Transfer transfer;
targetCount++;
transfer.numBytes = numBytesPerTransfer;
for (int x = 0; x < numSrcs; x++) transfer.srcs.push_back({memType, i});
// When using multiple destinations, the additional destinations are "local"
if (numDsts) transfer.dsts.push_back({memType, j});
for (int x = 1; x < numDsts; x++) transfer.dsts.push_back({memType, i});
transfer.exeDevice = {exeType, (useRemoteRead ? j : i)};
transfer.exeSubIndex = -1;
transfers.push_back(transfer);
}
}
} else {
// Each CU will work on all targets
for (int i = 0; i < numGpus; i++) {
TransferBench::Transfer transfer;
transfer.numBytes = numBytesPerTransfer;
transfer.exeDevice = {exeType, i};
transfer.exeSubIndex = -1;
targetCount = 0;
for (int j = 0; j < numGpus; j++) {
// Check whether or not to transfer to this GPU
if (i == j) {
if (!a2aLocal) continue;
} else if (a2aDirect) {
#if !defined(__NVCC__)
uint32_t linkType, hopCount;
HIP_CALL(hipExtGetLinkTypeAndHopCount(i, j, &linkType, &hopCount));
if (hopCount != 1) continue;
#endif
}
targetCount++;
for (int x = 0; x < numSrcs; x++) transfer.srcs.push_back({memType, useRemoteRead ? j : i});
if (numDsts) transfer.dsts.push_back({memType, j});
for (int x = 1; x < numDsts; x++) transfer.dsts.push_back({memType, i});
}
transfers.push_back(transfer);
}
}
printf("GPU-GFX All-To-All Sweep benchmark:\n");
printf("==========================\n");
printf("- Copying %lu bytes between %s pairs of GPUs\n", numBytesPerTransfer, a2aDirect ? "directly connected" : "all");
if (transfers.size() == 0) {
printf("[WARN} No transfers requested. Try adjusting A2A_DIRECT or A2A_LOCAL\n");
return;
}
// Execute Transfers
TransferBench::ConfigOptions cfg = ev.ToConfigOptions();
// Run tests
std::map<std::pair<int, int>, TransferBench::TestResults> results;
// Display summary
printf("#CUs\\Unroll");
for (int u : unrollList) {
printf(" %d(Min) ", u);
printf(" %d(Max) ", u);
}
printf("\n");
for (int c : numCusList) {
printf(" %5d ", c); fflush(stdout);
for (int u : unrollList) {
ev.gfxUnroll = cfg.gfx.unrollFactor = u;
for (auto& transfer : transfers)
transfer.numSubExecs = useSpray ? (c * targetCount) : c;
double minBandwidth = std::numeric_limits<double>::max();
double maxBandwidth = std::numeric_limits<double>::min();
TransferBench::TestResults result;
if (TransferBench::RunTransfers(cfg, transfers, result)) {
for (auto const& exeResult : result.exeResults) {
minBandwidth = std::min(minBandwidth, exeResult.second.avgBandwidthGbPerSec);
maxBandwidth = std::max(maxBandwidth, exeResult.second.avgBandwidthGbPerSec);
}
if (useSpray) {
minBandwidth *= targetCount;
maxBandwidth *= targetCount;
}
results[std::make_pair(c,u)] = result;
} else {
minBandwidth = 0.0;
}
printf(" %7.2f %7.2f ", minBandwidth, maxBandwidth); fflush(stdout);
}
printf("\n"); fflush(stdout);
}
if (verbose) {
int testNum = 0;
for (int c : numCusList) {
for (int u : unrollList) {
printf("CUs: %d Unroll %d\n", c, u);
PrintResults(ev, ++testNum, transfers, results[std::make_pair(c,u)]);
}
}
}
}
src/client/Presets/Presets.hpp
View file @ cd80b3a3
......@@ -24,6 +24,8 @@ THE SOFTWARE.
// Included after EnvVars and Executors
#include "AllToAll.hpp"
#include "AllToAllN.hpp"
#include "AllToAllSweep.hpp"
#include "HealthCheck.hpp"
#include "OneToAll.hpp"
#include "PeerToPeer.hpp"
......@@ -39,9 +41,11 @@ typedef void (*PresetFunc)(EnvVars& ev,
std::map<std::string, std::pair<PresetFunc, std::string>> presetFuncMap =
{
{"a2a", {AllToAllPreset, "Tests parallel transfers between all pairs of GPU devices"}},
{"healthcheck", {HealthCheckPreset,"Simple bandwidth health check (MI300X series only)"}},
{"a2a_n", {AllToAllRdmaPreset, "Tests parallel transfers between all pairs of GPU devices using Nearest NIC RDMA transfers"}},
{"a2asweep", {AllToAllSweepPreset, "Test GFX-based all-to-all transfers swept across different CU and GFX unroll counts"}},
{"healthcheck", {HealthCheckPreset, "Simple bandwidth health check (MI300X series only)"}},
{"one2all", {OneToAllPreset, "Test all subsets of parallel transfers from one GPU to all others"}},
{"p2p" , {PeerToPeerPreset, "Peer-to-peer device memory bandwidth test"}},
{"p2p" , {PeerToPeerPreset, " Peer-to-peer device memory bandwidth test"}},
{"rsweep", {SweepPreset, "Randomly sweep through sets of Transfers"}},
{"scaling", {ScalingPreset, "Run scaling test from one GPU to other devices"}},
{"schmoo", {SchmooPreset, "Scaling tests for local/remote read/write/copy"}},
......
src/client/Topology.hpp
View file @ cd80b3a3
......@@ -41,9 +41,9 @@ static int RemappedCpuIndex(int origIdx)
static void PrintNicToGPUTopo(bool outputToCsv)
{
#ifdef NIC_EXEC_ENABLED
printf(" NIC | Device Name | Active | PCIe Bus ID | NUMA | Closest GPU(s)\n");
printf(" NIC | Device Name | Active | PCIe Bus ID | NUMA | Closest GPU(s) | GID Index | GID Descriptor\n");
if(!outputToCsv)
printf("-----+-------------+--------+--------------+------+---------------\n");
printf("-----+-------------+--------+--------------+------+----------------+-----------+-------------------\n");
int numGpus = TransferBench::GetNumExecutors(EXE_GPU_GFX);
auto const& ibvDeviceList = GetIbvDeviceList();
......@@ -57,12 +57,15 @@ static void PrintNicToGPUTopo(bool outputToCsv)
}
}
printf(" %-3d | %-11s | %-6s | %-12s | %-4d | %-20s\n",
printf(" %-3d | %-11s | %-6s | %-12s | %-4d | %-14s | %-9s | %-20s\n",
i, ibvDeviceList[i].name.c_str(),
ibvDeviceList[i].hasActivePort ? "Yes" : "No",
ibvDeviceList[i].busId.c_str(),
ibvDeviceList[i].numaNode,
closestGpusStr.c_str());
closestGpusStr.c_str(),
ibvDeviceList[i].isRoce && ibvDeviceList[i].hasActivePort? std::to_string(ibvDeviceList[i].gidIndex).c_str() : "N/A",
ibvDeviceList[i].isRoce && ibvDeviceList[i].hasActivePort? ibvDeviceList[i].gidDescriptor.c_str() : "N/A"
);
}
printf("\n");
#endif
......
src/header/TransferBench.hpp
View file @ cd80b3a3
This diff is collapsed.
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment