From ae3dededed1c6c1a9b4c631d9edba690c4abb59c Mon Sep 17 00:00:00 2001
From: Yann Collet <cyan@fb.com>
Date: Tue, 20 Feb 2018 13:05:22 -0800
Subject: ensure bench speed measurement is more accurate for small inputs
Previous method would produce too many time() invocations,
becoming a significant fraction of workload measured.
The new strategy is to use time() only once per batch,
and dynamically resize batch size so that each round lasts approximately 1 second.
This only matters for small inputs.
Measurement for large files (such as silesia.tar) are much less impacted
(though decoding speed is so fast that even medium-size files will notice an improvement).
---
programs/bench.c | 57 +++++++++++++++++++++++++++++++++-----------------------
programs/util.h | 8 +++++++-
2 files changed, 41 insertions(+), 24 deletions(-)
diff --git a/programs/bench.c b/programs/bench.c
index 91292bc..0b1cf5c 100644
--- a/programs/bench.c
+++ b/programs/bench.c
@@ -67,6 +67,7 @@ static int LZ4_compress_local(const char* src, char* dst, int srcSize, int dstSi
#define NBSECONDS 3
#define TIMELOOP_MICROSEC 1*1000000ULL /* 1 second */
+#define TIMELOOP_NANOSEC 1*1000000000ULL /* 1 second */
#define ACTIVEPERIOD_MICROSEC 70*1000000ULL /* 70 seconds */
#define COOLPERIOD_SEC 10
#define DECOMP_MULT 1 /* test decompression DECOMP_MULT times longer than compression */
@@ -219,8 +220,9 @@ static int BMK_benchMem(const void* srcBuffer, size_t srcSize,
{ U64 fastestC = (U64)(-1LL), fastestD = (U64)(-1LL);
U64 const crcOrig = XXH64(srcBuffer, srcSize, 0);
UTIL_time_t coolTime;
- U64 const maxTime = (g_nbSeconds * TIMELOOP_MICROSEC) + 100;
- U32 nbDecodeLoops = ((200 MB) / (srcSize+1)) + 1; /* conservative decode speed estimate */
+ U64 const maxTime = (g_nbSeconds * TIMELOOP_NANOSEC) + 100;
+ U32 nbCompressionLoops = ((5 MB) / (srcSize+1)) + 1; /* conservative initial compression speed estimate */
+ U32 nbDecodeLoops = ((200 MB) / (srcSize+1)) + 1; /* conservative initial decode speed estimate */
U64 totalCTime=0, totalDTime=0;
U32 cCompleted=0, dCompleted=0;
# define NB_MARKS 4
@@ -232,9 +234,6 @@ static int BMK_benchMem(const void* srcBuffer, size_t srcSize,
coolTime = UTIL_getTime();
DISPLAYLEVEL(2, "\r%79s\r", "");
while (!cCompleted || !dCompleted) {
- UTIL_time_t clockStart;
- U64 clockLoop = g_nbSeconds ? TIMELOOP_MICROSEC : 1;
-
/* overheat protection */
if (UTIL_clockSpanMicro(coolTime) > ACTIVEPERIOD_MICROSEC) {
DISPLAYLEVEL(2, "\rcooling down ... \r");
@@ -248,21 +247,27 @@ static int BMK_benchMem(const void* srcBuffer, size_t srcSize,
UTIL_sleepMilli(1); /* give processor time to other processes */
UTIL_waitForNextTick();
- clockStart = UTIL_getTime();
if (!cCompleted) { /* still some time to do compression tests */
- U32 nbLoops = 0;
- do {
+ UTIL_time_t const clockStart = UTIL_getTime();
+ U32 nbLoops;
+ for (nbLoops=0; nbLoops < nbCompressionLoops; nbLoops++) {
U32 blockNb;
for (blockNb=0; blockNb<nbBlocks; blockNb++) {
size_t const rSize = compP.compressionFunction(blockTable[blockNb].srcPtr, blockTable[blockNb].cPtr, (int)blockTable[blockNb].srcSize, (int)blockTable[blockNb].cRoom, cLevel);
if (LZ4_isError(rSize)) EXM_THROW(1, "LZ4_compress() failed");
blockTable[blockNb].cSize = rSize;
+ } }
+ { U64 const clockSpan = UTIL_clockSpanNano(clockStart);
+ if (clockSpan > 0) {
+ if (clockSpan < fastestC * nbCompressionLoops)
+ fastestC = clockSpan / nbCompressionLoops;
+ assert(fastestC > 0);
+ nbCompressionLoops = (U32)(1000000000/*1sec*/ / fastestC) + 1; /* aim for ~1sec */
+ } else {
+ assert(nbCompressionLoops < 40000000); /* avoid overflow */
+ nbCompressionLoops *= 100;
}
- nbLoops++;
- } while (UTIL_clockSpanMicro(clockStart) < clockLoop);
- { U64 const clockSpan = UTIL_clockSpanMicro(clockStart);
- if (clockSpan < fastestC*nbLoops) fastestC = clockSpan / nbLoops;
totalCTime += clockSpan;
cCompleted = totalCTime>maxTime;
} }
@@ -274,7 +279,7 @@ static int BMK_benchMem(const void* srcBuffer, size_t srcSize,
markNb = (markNb+1) % NB_MARKS;
DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->%10u (%5.3f),%6.1f MB/s\r",
marks[markNb], displayName, (U32)srcSize, (U32)cSize, ratio,
- (double)srcSize / fastestC );
+ ((double)srcSize / fastestC) * 1000 );
(void)fastestD; (void)crcOrig; /* unused when decompression disabled */
#if 1
@@ -285,22 +290,28 @@ static int BMK_benchMem(const void* srcBuffer, size_t srcSize,
UTIL_waitForNextTick();
if (!dCompleted) {
+ UTIL_time_t const clockStart = UTIL_getTime();
U32 nbLoops;
- clockStart = UTIL_getTime();
for (nbLoops=0; nbLoops < nbDecodeLoops; nbLoops++) {
U32 blockNb;
for (blockNb=0; blockNb<nbBlocks; blockNb++) {
size_t const regenSize = LZ4_decompress_safe(blockTable[blockNb].cPtr, blockTable[blockNb].resPtr, (int)blockTable[blockNb].cSize, (int)blockTable[blockNb].srcSize);
if (LZ4_isError(regenSize)) {
- DISPLAY("LZ4_decompress_safe() failed on block %u \n", blockNb);
- clockLoop = 0; /* force immediate test end */
+ DISPLAY("LZ4_decompress_safe() failed on block %u \n", blockNb);
break;
}
blockTable[blockNb].resSize = regenSize;
} }
- { U64 const clockSpan = UTIL_clockSpanMicro(clockStart);
- if (clockSpan < fastestD*nbLoops) fastestD = clockSpan / nbLoops;
- nbDecodeLoops = (U32)(1000000/*1sec*/ / fastestD) + 1; /* aim for ~1sec */
+ { U64 const clockSpan = UTIL_clockSpanNano(clockStart);
+ if (clockSpan > 0) {
+ if (clockSpan < fastestD * nbDecodeLoops)
+ fastestD = clockSpan / nbDecodeLoops;
+ assert(fastestD > 0);
+ nbDecodeLoops = (U32)(1000000000/*1sec*/ / fastestD) + 1; /* aim for ~1sec */
+ } else {
+ assert(nbDecodeLoops < 40000000); /* avoid overflow */
+ nbDecodeLoops *= 100;
+ }
totalDTime += clockSpan;
dCompleted = totalDTime > (DECOMP_MULT*maxTime);
} }
@@ -308,8 +319,8 @@ static int BMK_benchMem(const void* srcBuffer, size_t srcSize,
markNb = (markNb+1) % NB_MARKS;
DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->%10u (%5.3f),%6.1f MB/s ,%6.1f MB/s\r",
marks[markNb], displayName, (U32)srcSize, (U32)cSize, ratio,
- (double)srcSize / fastestC,
- (double)srcSize / fastestD );
+ ((double)srcSize / fastestC) * 1000,
+ ((double)srcSize / fastestD) * 1000);
/* CRC Checking */
{ U64 const crcCheck = XXH64(resultBuffer, srcSize, 0);
@@ -339,8 +350,8 @@ static int BMK_benchMem(const void* srcBuffer, size_t srcSize,
} /* for (testNb = 1; testNb <= (g_nbSeconds + !g_nbSeconds); testNb++) */
if (g_displayLevel == 1) {
- double cSpeed = (double)srcSize / fastestC;
- double dSpeed = (double)srcSize / fastestD;
+ double const cSpeed = ((double)srcSize / fastestC) * 1000;
+ double const dSpeed = ((double)srcSize / fastestD) * 1000;
if (g_additionalParam)
DISPLAY("-%-3i%11i (%5.3f) %6.2f MB/s %6.1f MB/s %s (param=%d)\n", cLevel, (int)cSize, ratio, cSpeed, dSpeed, displayName, g_additionalParam);
else
diff --git a/programs/util.h b/programs/util.h
index a3576d7..ff25106 100644
--- a/programs/util.h
+++ b/programs/util.h
@@ -180,7 +180,7 @@ extern "C" {
mach_timebase_info(&rate);
init = 1;
}
- return (((clockEnd - clockStart) * (U64)rate.numer) / ((U64)rate.denom))/1000ULL;
+ return (((clockEnd - clockStart) * (U64)rate.numer) / ((U64)rate.denom)) / 1000ULL;
}
UTIL_STATIC U64 UTIL_getSpanTimeNano(UTIL_time_t clockStart, UTIL_time_t clockEnd)
{
@@ -249,6 +249,12 @@ UTIL_STATIC U64 UTIL_clockSpanMicro(UTIL_time_t clockStart)
return UTIL_getSpanTimeMicro(clockStart, clockEnd);
}
+/* returns time span in nanoseconds */
+UTIL_STATIC U64 UTIL_clockSpanNano(UTIL_time_t clockStart)
+{
+ UTIL_time_t const clockEnd = UTIL_getTime();
+ return UTIL_getSpanTimeNano(clockStart, clockEnd);
+}
UTIL_STATIC void UTIL_waitForNextTick(void)
{
--
cgit v0.12