Skip to content

Improve GPU memory benchmark #7495

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
davidrohr merged 8 commits into from
Nov 3, 2021
Merged

Improve GPU memory benchmark #7495

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
7170097
Add boundaries protection + float
mconcas Nov 2, 2021
6197883
Add Host-based throughput computation
mconcas Nov 2, 2021
337f682
Fix boundary checking bug
mconcas Nov 2, 2021
63700ab
Add proto function for distributed benchmark
mconcas Nov 2, 2021
cbbe3e4
Add read distributed kernel (segfaults)
mconcas Nov 3, 2021
d7fba08
Fix read distributed kernel
mconcas Nov 3, 2021
271919e
Compute throughput for distributed case
mconcas Nov 3, 2021
3bbedec
Add rest of the kernels
mconcas Nov 3, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 13 additions & 4 deletions GPU/GPUbenchmark/Shared/Kernels.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,25 +39,34 @@ class GPUbenchmark final
template <typename... T>
float measure(void (GPUbenchmark::*)(T...), const char*, T&&... args);

// Single stream synchronous (sequential kernels) execution
// Single stream (sequential kernels) execution
template <typename... T>
float runSequential(void (*kernel)(chunk_t*, size_t, T...),
std::pair<int, int>& chunkRanges,
std::pair<float, float>& chunkRanges,
int nLaunches,
int dimGrid,
int dimBlock,
T&... args);

// Multi-streams asynchronous executions on whole memory
// Multi-streams asynchronous executions
template <typename... T>
std::vector<float> runConcurrent(void (*kernel)(chunk_t*, size_t, T...),
std::vector<std::pair<int, int>>& chunkRanges,
std::vector<std::pair<float, float>>& chunkRanges,
int nLaunches,
int dimStreams,
int nBlocks,
int nThreads,
T&... args);

// Single stream executions on all chunks at a time by same kernel
template <typename... T>
float runDistributed(void (*kernel)(chunk_t**, size_t*, T...),
std::vector<std::pair<float, float>>& chunkRanges,
int nLaunches,
int nBlocks,
int nThreads,
T&... args);

// Main interface
void globalInit(); // Allocate scratch buffers and compute runtime parameters
void run(); // Execute all specified callbacks
Expand Down
30 changes: 14 additions & 16 deletions GPU/GPUbenchmark/Shared/Utils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,7 +72,8 @@ inline std::ostream& operator<<(std::ostream& os, Test test)

enum class Mode {
Sequential,
Concurrent
Concurrent,
Distributed
};

inline std::ostream& operator<<(std::ostream& os, Mode mode)
Expand All @@ -84,6 +85,9 @@ inline std::ostream& operator<<(std::ostream& os, Mode mode)
case Mode::Concurrent:
os << "concurrent";
break;
case Mode::Distributed:
os << "distributed";
break;
}
return os;
}
Expand Down Expand Up @@ -138,17 +142,11 @@ inline std::string getTestName(Mode mode, Test test, KernelConfig blocks)
return tname;
}

template <class chunk_t>
inline chunk_t* getPartPtr(chunk_t* scratchPtr, float chunkReservedGB, int partNumber)
{
return reinterpret_cast<chunk_t*>(reinterpret_cast<char*>(scratchPtr) + static_cast<size_t>(GB * chunkReservedGB) * partNumber);
}

// Return pointer to custom offset (GB)
template <class chunk_t>
inline chunk_t* getCustomPtr(chunk_t* scratchPtr, int partNumber)
inline chunk_t* getCustomPtr(chunk_t* scratchPtr, float startGB)
{
return reinterpret_cast<chunk_t*>(reinterpret_cast<char*>(scratchPtr) + static_cast<size_t>(GB * partNumber));
return reinterpret_cast<chunk_t*>(reinterpret_cast<char*>(scratchPtr) + static_cast<size_t>(GB * startGB));
}

inline float computeThroughput(Test test, float result, float chunkSizeGB, int ntests)
Expand All @@ -160,9 +158,9 @@ inline float computeThroughput(Test test, float result, float chunkSizeGB, int n
}

template <class chunk_t>
inline size_t getBufferCapacity(int chunkReservedGB)
inline size_t getBufferCapacity(float chunkReservedGB)
{
return static_cast<size_t>(GB * chunkReservedGB / sizeof(chunk_t));
return static_cast<size_t>((GB * chunkReservedGB) / sizeof(chunk_t));
}

// LCG: https://rosettacode.org/wiki/Linear_congruential_generator
Expand Down Expand Up @@ -202,7 +200,7 @@ struct benchmarkOpts {
std::vector<Mode> modes = {Mode::Sequential, Mode::Concurrent};
std::vector<KernelConfig> pools = {KernelConfig::Single, KernelConfig::Multi};
std::vector<std::string> dtypes = {"char", "int", "ulong"};
std::vector<std::pair<int, int>> testChunks;
std::vector<std::pair<float, float>> testChunks;
float chunkReservedGB = 1.f;
float threadPoolFraction = 1.f;
float freeMemoryFractionToAllocate = 0.95f;
Expand Down Expand Up @@ -235,10 +233,10 @@ struct gpuState {
float chunkReservedGB; // Size of each partition (GB)

// General containers and state
chunk_t* scratchPtr; // Pointer to scratch buffer
size_t scratchSize; // Size of scratch area (B)
std::vector<chunk_t*> partAddrOnHost; // Pointers to scratch partitions on host vector
std::vector<std::pair<int, int>> testChunks; // Vector of definitions for arbitrary chunks
chunk_t* scratchPtr; // Pointer to scratch buffer
size_t scratchSize; // Size of scratch area (B)
std::vector<chunk_t*> partAddrOnHost; // Pointers to scratch partitions on host vector
std::vector<std::pair<float, float>> testChunks; // Vector of definitions for arbitrary chunks

// Static info
size_t totalMemory;
Expand Down
Loading