Skip to content

Fix the heuristics for L3 cache size for Arm64 #71029

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
kunalspathak merged 13 commits into from
Jun 22, 2022
Merged

Fix the heuristics for L3 cache size for Arm64 #71029

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
13 commits
Select commit Hold shift + click to select a range
f6310e4
Get cacheLevel for unix
kunalspathak Jun 20, 2022
45c4138
Get cacheLevel for windows
kunalspathak Jun 20, 2022
8b6862c
Get cacheLevel for os
kunalspathak Jun 20, 2022
363490c
add the macro CHECK_CACHE_SIZE
kunalspathak Jun 20, 2022
29ed6db
fix multiplier
kunalspathak Jun 21, 2022
468dc4a
Review comments
kunalspathak Jun 21, 2022
b09e495
Fix linux arm64 build
kunalspathak Jun 21, 2022
c7c905a
Change from std:string to char[]
kunalspathak Jun 21, 2022
0bf5ec1
minor nit
kunalspathak Jun 21, 2022
1f7ce1c
opps should be outside the loop
kunalspathak Jun 21, 2022
cc72139
Replace assert with _ASSERTE
kunalspathak Jun 21, 2022
5671866
CHECK_CACHE_SIZE -> UPDATE_CACHE_SIZE_AND_LEVEL
kunalspathak Jun 21, 2022
aaab0bf
Use GetTotalProcessorCount for windows
kunalspathak Jun 22, 2022
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
Prev Previous commit
Next Next commit
Review comments
  • Loading branch information
@kunalspathak
kunalspathak committed Jun 21, 2022
commit 468dc4a479b74e580abfc1ea02f68e672537842a
62 changes: 38 additions & 24 deletions src/coreclr/gc/unix/gcenv.unix.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -929,7 +929,7 @@ static size_t GetLogicalProcessorCacheSizeFromOS()
#endif

#if (defined(HOST_ARM64) || defined(HOST_LOONGARCH64)) && !defined(TARGET_OSX)
if ((cacheSize == 0) || (cacheLevel != 3))
if (cacheSize == 0)
{
// We expect to get the L3 cache size for Arm64 but currently expected to be missing that info
// from most of the machines.
Expand All @@ -947,30 +947,12 @@ static size_t GetLogicalProcessorCacheSizeFromOS()
// If we use recent high core count chips as a guide for state of the art, we find
// total L3 cache to be 1-2MB/core. As always, there are exceptions.

// Hence, just use the following heuristics at best depending on the CPU count
// 1 ~ 4 : 4 MB
// 5 ~ 16 : 8 MB
// 17 ~ 64 : 16 MB
// 65+ : 32 MB
// Estimate cache size based on CPU count
// Assume lower core count are lighter weight parts which are likely to have smaller caches
// Assume L3$/CPU grows linearly from 256K to 1.5M/CPU as logicalCPUs grows from 2 to 12 CPUs
DWORD logicalCPUs = g_totalCpuCount;
if (logicalCPUs < 5)
{
cacheSize = 4;
}
else if (logicalCPUs < 17)
{
cacheSize = 8;
}
else if (logicalCPUs < 65)
{
cacheSize = 16;
}
else
{
cacheSize = 32;
}

cacheSize = cacheSize * 1024;
cacheSize = logicalCPUs * std::min(1536, std::max(256, (int)logicalCPUs * 128)) * 1024;
}
#endif

Expand All @@ -995,6 +977,38 @@ static size_t GetLogicalProcessorCacheSizeFromOS()
}
#endif

#if (defined(HOST_ARM64) || defined(HOST_LOONGARCH64)) && !defined(TARGET_OSX)
if (cacheLevel != 3)
{
// We expect to get the L3 cache size for Arm64 but currently expected to be missing that info
// from most of the machines.
// Hence, just use the following heuristics at best depending on the CPU count
// 1 ~ 4 : 4 MB
// 5 ~ 16 : 8 MB
// 17 ~ 64 : 16 MB
// 65+ : 32 MB
DWORD logicalCPUs = g_totalCpuCount;
if (logicalCPUs < 5)
{
cacheSize = 4;
}
else if (logicalCPUs < 17)
{
cacheSize = 8;
}
else if (logicalCPUs < 65)
{
cacheSize = 16;
}
else
{
cacheSize = 32;
}

cacheSize *= (1024 * 1024);
}
#endif

return cacheSize;
}

Expand Down Expand Up @@ -1071,7 +1085,7 @@ size_t GCToOSInterface::GetCacheSizePerLogicalCpu(bool trueSize)
s_maxSize = maxSize;
s_maxTrueSize = maxTrueSize;

// printf("GetCacheSizePerLogicalCpu returns %d, adjusted size %d\n", maxSize, maxTrueSize);
// printf("GetCacheSizePerLogicalCpu returns %zu, adjusted size %zu\n", maxSize, maxTrueSize);
return trueSize ? maxTrueSize : maxSize;
}

Expand Down
24 changes: 14 additions & 10 deletions src/coreclr/gc/windows/gcenv.windows.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -403,6 +403,7 @@ size_t GetLogicalProcessorCacheSizeFromOS()
{
size_t cache_size = 0;
size_t cache_level = 0;
uint32_t totalCPUCount = 0;

DWORD nEntries = 0;

Expand Down Expand Up @@ -432,6 +433,10 @@ size_t GetLogicalProcessorCacheSizeFromOS()
cache_level = pslpi[i].Cache.Level;
}
}
else if (pslpi[i].Relationship == RelationProcessorCore)
{
totalCPUCount++;
}
}
cache_size = last_cache_size;
}
Expand All @@ -450,25 +455,24 @@ size_t GetLogicalProcessorCacheSizeFromOS()
// 5 ~ 16 : 8 MB
// 17 ~ 64 : 16 MB
// 65+ : 32 MB
uint32_t logicalCPUs = GetTotalProcessorCount();
if (logicalCPUs < 5)
if (totalCPUCount < 5)
{
cacheSize = 4;
cache_size = 4;
}
else if (logicalCPUs < 17)
else if (totalCPUCount < 17)
{
cacheSize = 8;
cache_size = 8;
}
else if (logicalCPUs < 65)
else if (totalCPUCount < 65)
{
cacheSize = 16;
cache_size = 16;
}
else
{
cacheSize = 32;
cache_size = 32;
}

cacheSize *= 1024;
cache_size *= (1024 * 1024);
}
#endif // TARGET_ARM64

Expand Down Expand Up @@ -877,7 +881,7 @@ size_t GCToOSInterface::GetCacheSizePerLogicalCpu(bool trueSize)
s_maxSize = maxSize;
s_maxTrueSize = maxTrueSize;

// printf("GetCacheSizePerLogicalCpu returns %d, adjusted size %d\n", maxSize, maxTrueSize);
// printf("GetCacheSizePerLogicalCpu returns %zu, adjusted size %zu\n", maxSize, maxTrueSize);
return trueSize ? maxTrueSize : maxSize;
}

Expand Down
60 changes: 37 additions & 23 deletions src/coreclr/pal/src/misc/sysinfo.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -594,7 +594,7 @@ PAL_GetLogicalProcessorCacheSizeFromOS()
#endif

#if (defined(HOST_ARM64) || defined(HOST_LOONGARCH64)) && !defined(TARGET_OSX)
if ((cacheSize == 0) || (cacheLevel != 3))
if (cacheSize == 0)
{
// We expect to get the L3 cache size for Arm64 but currently expected to be missing that info
// from most of the machines with an exceptions on some machines.
Expand All @@ -612,30 +612,12 @@ PAL_GetLogicalProcessorCacheSizeFromOS()
// If we use recent high core count chips as a guide for state of the art, we find
// total L3 cache to be 1-2MB/core. As always, there are exceptions.

// Hence, just use the following heuristics at best depending on the CPU count
// 1 ~ 4 : 4 MB
// 5 ~ 16 : 8 MB
// 17 ~ 64 : 16 MB
// 65+ : 32 MB
// Estimate cache size based on CPU count
// Assume lower core count are lighter weight parts which are likely to have smaller caches
// Assume L3$/CPU grows linearly from 256K to 1.5M/CPU as logicalCPUs grows from 2 to 12 CPUs
DWORD logicalCPUs = PAL_GetLogicalCpuCountFromOS();
if (logicalCPUs < 5)
{
cacheSize = 4;
}
else if (logicalCPUs < 17)
{
cacheSize = 8;
}
else if (logicalCPUs < 65)
{
cacheSize = 16;
}
else
{
cacheSize = 32;
}

cacheSize = cacheSize * 1024;
cacheSize = logicalCPUs*std::min(1536, std::max(256, (int)logicalCPUs*128))*1024;
}
#endif

Expand All @@ -660,5 +642,37 @@ PAL_GetLogicalProcessorCacheSizeFromOS()
}
#endif

#if (defined(HOST_ARM64) || defined(HOST_LOONGARCH64)) && !defined(TARGET_OSX)
if (cacheLevel != 3)
{
// We expect to get the L3 cache size for Arm64 but currently expected to be missing that info
// from most of the machines.
// Hence, just use the following heuristics at best depending on the CPU count
// 1 ~ 4 : 4 MB
// 5 ~ 16 : 8 MB
// 17 ~ 64 : 16 MB
// 65+ : 32 MB
DWORD logicalCPUs = g_totalCpuCount;
if (logicalCPUs < 5)
{
cacheSize = 4;
}
else if (logicalCPUs < 17)
{
cacheSize = 8;
}
else if (logicalCPUs < 65)
{
cacheSize = 16;
}
else
{
cacheSize = 32;
}

cacheSize *= (1024 * 1024);
}
#endif

return cacheSize;
}
33 changes: 20 additions & 13 deletions src/coreclr/vm/gcenv.os.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -589,6 +589,7 @@ size_t GetLogicalProcessorCacheSizeFromOS()
{
size_t cache_size = 0;
size_t cache_level = 0;
uint32_t totalCPUCount = 0;
DWORD nEntries = 0;

// Try to use GetLogicalProcessorInformation API and get a valid pointer to the SLPI array if successful. Returns NULL
Expand All @@ -609,10 +610,17 @@ size_t GetLogicalProcessorCacheSizeFromOS()

for (DWORD i=0; i < nEntries; i++)
{
if (last_cache_size < pslpi[i].Cache.Size)
if (pslpi[i].Relationship == RelationCache)
{
last_cache_size = pslpi[i].Cache.Size;
cache_level = pslpi[i].Cache.Level;
if (last_cache_size < pslpi[i].Cache.Size)
{
last_cache_size = pslpi[i].Cache.Size;
cache_level = pslpi[i].Cache.Level;
}
}
else if (pslpi[i].Relationship == RelationProcessorCore)
{
totalCPUCount++;
}
}
cache_size = last_cache_size;
Expand All @@ -632,25 +640,24 @@ size_t GetLogicalProcessorCacheSizeFromOS()
// 5 ~ 16 : 8 MB
// 17 ~ 64 : 16 MB
// 65+ : 32 MB
uint32_t logicalCPUs = GetTotalProcessorCount();
if (logicalCPUs < 5)
if (totalCPUCount < 5)
{
cacheSize = 4;
cache_size = 4;
}
else if (logicalCPUs < 17)
else if (totalCPUCount < 17)
{
cacheSize = 8;
cache_size = 8;
}
else if (logicalCPUs < 65)
else if (totalCPUCount < 65)
{
cacheSize = 16;
cache_size = 16;
}
else
{
cacheSize = 32;
cache_size = 32;
}

cacheSize *= 1024;
cache_size *= (1024 * 1024);
}
#endif // TARGET_ARM64

Expand Down Expand Up @@ -683,7 +690,7 @@ size_t GCToOSInterface::GetCacheSizePerLogicalCpu(bool trueSize)
s_maxSize = maxSize;
s_maxTrueSize = maxTrueSize;

// printf("GetCacheSizePerLogicalCpu returns %d, adjusted size %d\n", maxSize, maxTrueSize);
// printf("GetCacheSizePerLogicalCpu returns %zu, adjusted size %zu\n", maxSize, maxTrueSize);
return trueSize ? maxTrueSize : maxSize;
}

Expand Down