From: Ben Peart Date: Sat, 10 Jun 2017 14:04:29 GMT Subject: Re: [PATCH v5 7/7] fsmonitor: add a performance test Message-ID: <5d0d07a8-c47c-e596-5ea2-4ca1e041f77a@gmail.com> In-Reply-To: <20170610134026.104552-8-benpeart@microsoft.com> Here are some perf test results for repos of various size generated with many-files.sh. Comparing cold fs cache times on a fast SSD running Windows: # files preloadindex fsmonitor reduction ========================================================= 10,000 0.69 0.46 33% 100,000 3.75 0.7 81% 1,000,000 35.07 3.24 91% 10,000 files GIT_PERF_REPEAT_COUNT=1 GIT_PERF_LARGE_REPO=/c/Repos/gen-many-files-3.10.9.git/ ./run p7519-fsmonitor.sh Test this tree --------------------------------------------------------------------------------- 7519.3: status (fsmonitor=false, cold fs cache) 0.69(0.03+0.06) 7519.4: status (fsmonitor=false, warm fs cache) 0.45(0.01+0.06) 7519.6: status -uno (fsmonitor=false, cold fs cache) 0.53(0.03+0.04) 7519.8: status -uall (fsmonitor=false, cold fs cache) 0.57(0.04+0.01) 7519.11: status (fsmonitor=true, cold fs cache, cold fsmonitor) 2.16(0.01+0.06) 7519.12: status (fsmonitor=true, warm fs cache, warm fsmonitor) 0.36(0.01+0.07) 7519.14: status (fsmonitor=true, cold fs cache, warm fsmonitor) 0.46(0.04+0.06) 7519.16: status -uno (fsmonitor=true, cold fs cache) 0.46(0.03+0.04) 7519.18: status -uall (fsmonitor=true, cold fs cache) 0.76(0.04+0.04) 100,000 files GIT_PERF_REPEAT_COUNT=1 GIT_PERF_LARGE_REPO=/c/Repos/gen-many-files-4.10.9.git/ ./run p7519-fsmonitor.sh Test this tree ---------------------------------------------------------------------------------- 7519.3: status (fsmonitor=false, cold fs cache) 3.75(0.01+0.04) 7519.4: status (fsmonitor=false, warm fs cache) 2.74(0.06+0.06) 7519.6: status -uno (fsmonitor=false, cold fs cache) 2.88(0.01+0.06) 7519.8: status -uall (fsmonitor=false, cold fs cache) 3.24(0.00+0.06) 7519.11: status (fsmonitor=true, cold fs cache, cold fsmonitor) 17.90(0.00+0.10) 7519.12: status (fsmonitor=true, warm fs cache, warm fsmonitor) 0.59(0.00+0.09) 7519.14: status (fsmonitor=true, cold fs cache, warm fsmonitor) 0.70(0.00+0.07) 7519.16: status -uno (fsmonitor=true, cold fs cache) 0.69(0.01+0.01) 7519.18: status -uall (fsmonitor=true, cold fs cache) 4.51(0.00+0.06) 1,000,000 files GIT_PERF_REPEAT_COUNT=1 GIT_PERF_LARGE_REPO=/c/Repos/gen-many-files-5.10.9.git/ ./run p7519-fsmonitor.sh Test this tree ----------------------------------------------------------------------------------- 7519.3: status (fsmonitor=false, cold fs cache) 35.07(0.01+0.03) 7519.4: status (fsmonitor=false, warm fs cache) 26.58(0.03+0.07) 7519.6: status -uno (fsmonitor=false, cold fs cache) 26.46(0.03+0.06) 7519.8: status -uall (fsmonitor=false, cold fs cache) 31.55(0.01+0.03) 7519.11: status (fsmonitor=true, cold fs cache, cold fsmonitor) 193.15(0.01+0.04) 7519.12: status (fsmonitor=true, warm fs cache, warm fsmonitor) 3.03(0.01+0.07) 7519.14: status (fsmonitor=true, cold fs cache, warm fsmonitor) 3.24(0.01+0.04) 7519.16: status -uno (fsmonitor=true, cold fs cache) 2.99(0.03+0.03) 7519.18: status -uall (fsmonitor=true, cold fs cache) 35.07(0.03+0.07) On 6/10/2017 9:40 AM, Ben Peart wrote: > Add a test utility (test-drop-caches) that enables dropping the file > system cache on Windows. > > Add a perf test (p7519-fsmonitor.sh) for fsmonitor. > > Signed-off-by: Ben Peart > Signed-off-by: Ævar Arnfjörð Bjarmason > --- > Makefile | 1 + > t/helper/test-drop-caches.c | 107 +++++++++++++++++++++++++++++ > t/perf/p7519-fsmonitor.sh | 161 ++++++++++++++++++++++++++++++++++++++++++++ > 3 files changed, 269 insertions(+) > create mode 100644 t/helper/test-drop-caches.c > create mode 100755 t/perf/p7519-fsmonitor.sh > > diff --git a/Makefile b/Makefile > index 992dd58801..893947839f 100644 > --- a/Makefile > +++ b/Makefile > @@ -648,6 +648,7 @@ TEST_PROGRAMS_NEED_X += test-subprocess > TEST_PROGRAMS_NEED_X += test-svn-fe > TEST_PROGRAMS_NEED_X += test-urlmatch-normalization > TEST_PROGRAMS_NEED_X += test-wildmatch > +TEST_PROGRAMS_NEED_X += test-drop-caches > > TEST_PROGRAMS = $(patsubst %,t/helper/%$X,$(TEST_PROGRAMS_NEED_X)) > > diff --git a/t/helper/test-drop-caches.c b/t/helper/test-drop-caches.c > new file mode 100644 > index 0000000000..80830d920b > --- /dev/null > +++ b/t/helper/test-drop-caches.c > @@ -0,0 +1,107 @@ > +#include "git-compat-util.h" > +#include > + > +typedef DWORD NTSTATUS; > + > +#ifdef GIT_WINDOWS_NATIVE > +#include > + > +#define STATUS_SUCCESS (0x00000000L) > +#define STATUS_PRIVILEGE_NOT_HELD (0xC0000061L) > + > +typedef enum _SYSTEM_INFORMATION_CLASS { > + SystemMemoryListInformation = 80, // 80, q: SYSTEM_MEMORY_LIST_INFORMATION; s: SYSTEM_MEMORY_LIST_COMMAND (requires SeProfileSingleProcessPrivilege) > +} SYSTEM_INFORMATION_CLASS; > + > +// private > +typedef enum _SYSTEM_MEMORY_LIST_COMMAND > +{ > + MemoryCaptureAccessedBits, > + MemoryCaptureAndResetAccessedBits, > + MemoryEmptyWorkingSets, > + MemoryFlushModifiedList, > + MemoryPurgeStandbyList, > + MemoryPurgeLowPriorityStandbyList, > + MemoryCommandMax > +} SYSTEM_MEMORY_LIST_COMMAND; > + > +BOOL GetPrivilege(HANDLE TokenHandle, LPCSTR lpName, int flags) > +{ > + BOOL bResult; > + DWORD dwBufferLength; > + LUID luid; > + TOKEN_PRIVILEGES tpPreviousState; > + TOKEN_PRIVILEGES tpNewState; > + > + dwBufferLength = 16; > + bResult = LookupPrivilegeValueA(0, lpName, &luid); > + if (bResult) > + { > + tpNewState.PrivilegeCount = 1; > + tpNewState.Privileges[0].Luid = luid; > + tpNewState.Privileges[0].Attributes = 0; > + bResult = AdjustTokenPrivileges(TokenHandle, 0, &tpNewState, (DWORD)((LPBYTE)&(tpNewState.Privileges[1]) - (LPBYTE)&tpNewState), &tpPreviousState, &dwBufferLength); > + if (bResult) > + { > + tpPreviousState.PrivilegeCount = 1; > + tpPreviousState.Privileges[0].Luid = luid; > + tpPreviousState.Privileges[0].Attributes = flags != 0 ? 2 : 0; > + bResult = AdjustTokenPrivileges(TokenHandle, 0, &tpPreviousState, dwBufferLength, 0, 0); > + } > + } > + return bResult; > +} > +#endif > + > +int cmd_main(int argc, const char **argv) > +{ > + NTSTATUS status = 1; > +#ifdef GIT_WINDOWS_NATIVE > + HANDLE hProcess = GetCurrentProcess(); > + HANDLE hToken; > + if (!OpenProcessToken(hProcess, TOKEN_QUERY | TOKEN_ADJUST_PRIVILEGES, &hToken)) > + { > + _ftprintf(stderr, _T("Can't open current process token\n")); > + return 1; > + } > + > + if (!GetPrivilege(hToken, "SeProfileSingleProcessPrivilege", 1)) > + { > + _ftprintf(stderr, _T("Can't get SeProfileSingleProcessPrivilege\n")); > + return 1; > + } > + > + CloseHandle(hToken); > + > + HMODULE ntdll = LoadLibrary(_T("ntdll.dll")); > + if (!ntdll) > + { > + _ftprintf(stderr, _T("Can't load ntdll.dll, wrong Windows version?\n")); > + return 1; > + } > + > + NTSTATUS(WINAPI *NtSetSystemInformation)(INT, PVOID, ULONG) = (NTSTATUS(WINAPI *)(INT, PVOID, ULONG))GetProcAddress(ntdll, "NtSetSystemInformation"); > + if (!NtSetSystemInformation) > + { > + _ftprintf(stderr, _T("Can't get function addresses, wrong Windows version?\n")); > + return 1; > + } > + > + SYSTEM_MEMORY_LIST_COMMAND command = MemoryPurgeStandbyList; > + status = NtSetSystemInformation( > + SystemMemoryListInformation, > + &command, > + sizeof(SYSTEM_MEMORY_LIST_COMMAND) > + ); > + if (status == STATUS_PRIVILEGE_NOT_HELD) > + { > + _ftprintf(stderr, _T("Insufficient privileges to execute the memory list command")); > + } > + else if (status != STATUS_SUCCESS) > + { > + _ftprintf(stderr, _T("Unable to execute the memory list command %lX"), status); > + } > +#endif > + > + return status; > +} > diff --git a/t/perf/p7519-fsmonitor.sh b/t/perf/p7519-fsmonitor.sh > new file mode 100755 > index 0000000000..e41905cb9b > --- /dev/null > +++ b/t/perf/p7519-fsmonitor.sh > @@ -0,0 +1,161 @@ > +#!/bin/sh > + > +test_description="Test core.fsmonitor" > + > +. ./perf-lib.sh > + > +# This has to be run with GIT_PERF_REPEAT_COUNT=1 to generate valid results. > +# Otherwise the caching that happens for the nth run will negate the validity > +# of the comparisons. > +if [ "$GIT_PERF_REPEAT_COUNT" -ne 1 ] > +then > + echo "warning: This test must be run with GIT_PERF_REPEAT_COUNT=1 to generate valid results." >&2 > + echo "warning: Setting GIT_PERF_REPEAT_COUNT=1" >&2 > + GIT_PERF_REPEAT_COUNT=1 > +fi > + > +test_perf_large_repo > +test_checkout_worktree > + > +# Convert unix style paths to what Watchman expects > +case "$(uname -s)" in > +MINGW*|MSYS_NT*) > + GIT_WORK_TREE="$(cygpath -aw "$PWD" | sed 's,\\,/,g')" > + ;; > +*) > + GIT_WORK_TREE="$PWD" > + ;; > +esac > + > +# The big win for using fsmonitor is the elimination of the need to scan > +# the working directory looking for changed files and untracked files. If > +# the file information is all cached in RAM, the benefits are reduced. > + > +flush_disk_cache () { > + case "$(uname -s)" in > + MINGW*|MSYS_NT*) > + sync && test-drop-caches > + ;; > + *) > + sudo sync && echo 3 | sudo tee /proc/sys/vm/drop_caches > + ;; > + esac > + > +} > + > +test_lazy_prereq UNTRACKED_CACHE ' > + { git update-index --test-untracked-cache; ret=$?; } && > + test $ret -ne 1 > +' > + > +test_expect_success "setup" ' > + # Maybe set untrackedCache & splitIndex depending on the environment > + if test -n "$GIT_PERF_7519_UNTRACKED_CACHE" > + then > + git config core.untrackedCache "$GIT_PERF_7519_UNTRACKED_CACHE" > + else > + if test_have_prereq UNTRACKED_CACHE > + then > + git config core.untrackedCache true > + else > + git config core.untrackedCache false > + fi > + fi && > + > + if test -n "$GIT_PERF_7519_SPLIT_INDEX" > + then > + git config core.splitIndex "$GIT_PERF_7519_SPLIT_INDEX" > + fi && > + > + # Hook scaffolding > + mkdir .git/hooks && > + cp ../../../templates/hooks--query-fsmonitor.sample .git/hooks/query-fsmonitor && > + > + # have Watchman monitor the test folder > + watchman watch "$GIT_WORK_TREE" && > + watchman watch-list | grep -q -F "$GIT_WORK_TREE" > +' > + > +# Worst case without fsmonitor > +test_expect_success "clear fs cache" ' > + git config core.fsmonitor false && > + flush_disk_cache > +' > +test_perf "status (fsmonitor=false, cold fs cache)" ' > + git status > +' > + > +# Best case without fsmonitor > +test_perf "status (fsmonitor=false, warm fs cache)" ' > + git status > +' > + > +# Let's see if -uno & -uall make any difference > +test_expect_success "clear fs cache" ' > + flush_disk_cache > +' > +test_perf "status -uno (fsmonitor=false, cold fs cache)" ' > + git status -uno > +' > + > +test_expect_success "clear fs cache" ' > + flush_disk_cache > +' > +test_perf "status -uall (fsmonitor=false, cold fs cache)" ' > + git status -uall > +' > + > +# The first run with core.fsmonitor=true has to do a normal scan and write > +# out the index extension. > +test_expect_success "populate extension" ' > + # core.preloadIndex defeats the benefits of core.fsMonitor as it > + # calls lstat for the index entries. Turn it off as _not_ doing > + # the work is faster than doing the work across multiple threads. > + git config core.fsmonitor true && > + git config core.preloadIndex false && > + git status > +' > + > +# Worst case with fsmonitor > +test_expect_success "shutdown fsmonitor, clear fs cache" ' > + watchman shutdown-server && > + flush_disk_cache > +' > +test_perf "status (fsmonitor=true, cold fs cache, cold fsmonitor)" ' > + git status > +' > + > +# Best case with fsmonitor > +test_perf "status (fsmonitor=true, warm fs cache, warm fsmonitor)" ' > + git status > +' > + > +# Best improved with fsmonitor (compare to worst case without fsmonitor) > +test_expect_success "clear fs cache" ' > + flush_disk_cache > +' > +test_perf "status (fsmonitor=true, cold fs cache, warm fsmonitor)" ' > + git status > +' > + > +# Let's see if -uno & -uall make any difference > +test_expect_success "clear fs cache" ' > + flush_disk_cache > +' > +test_perf "status -uno (fsmonitor=true, cold fs cache)" ' > + git status -uno > +' > + > +test_expect_success "clear fs cache" ' > + flush_disk_cache > +' > +test_perf "status -uall (fsmonitor=true, cold fs cache)" ' > + git status -uall > +' > + > +test_expect_success "cleanup" ' > + watchman watch-del "$GIT_WORK_TREE" && > + watchman shutdown-server > +' > + > +test_done >