Add option to use logical cores for default threads

diff --git a/programs/util.c b/programs/util.c
index 96459b9d39cc009b045670bbf31f0f0366b32ae0..81c10d0cae263c537c7e7d1a5e8dbe6a0fd4b26e 100644 (file)
--- a/programs/util.c
+++ b/programs/util.c
@@ -1103,7 +1103,7 @@ FileNamesTable* UTIL_createFNT_fromROTable(const char** filenames, size_t nbFile
 
 
 /*-****************************************
-*  count the number of physical cores
+*  count the number of cores
 ******************************************/
 
 #if defined(_WIN32) || defined(WIN32)
@@ -1112,10 +1112,26 @@ FileNamesTable* UTIL_createFNT_fromROTable(const char** filenames, size_t nbFile
 
 typedef BOOL(WINAPI* LPFN_GLPI)(PSYSTEM_LOGICAL_PROCESSOR_INFORMATION, PDWORD);
 
-int UTIL_countPhysicalCores(void)
+DWORD CountSetBits(ULONG_PTR bitMask)
+{
+    DWORD LSHIFT = sizeof(ULONG_PTR)*8 - 1;
+    DWORD bitSetCount = 0;
+    ULONG_PTR bitTest = (ULONG_PTR)1 << LSHIFT;    
+    DWORD i;
+    
+    for (i = 0; i <= LSHIFT; ++i)
+    {
+        bitSetCount += ((bitMask & bitTest)?1:0);
+        bitTest/=2;
+    }
+
+    return bitSetCount;
+}
+
+int UTIL_countCores(int logical)
 {
-    static int numPhysicalCores = 0;
-    if (numPhysicalCores != 0) return numPhysicalCores;
+    static int numCores = 0;
+    if (numCores != 0) return numCores;
 
     {   LPFN_GLPI glpi;
         BOOL done = FALSE;
@@ -1161,7 +1177,10 @@ int UTIL_countPhysicalCores(void)
         while (byteOffset + sizeof(SYSTEM_LOGICAL_PROCESSOR_INFORMATION) <= returnLength) {
 
             if (ptr->Relationship == RelationProcessorCore) {
-                numPhysicalCores++;
+                if (logical)
+                    numCores += CountSetBits(ptr->ProcessorMask);
+                else
+                    numCores++;
             }
 
             ptr++;
@@ -1170,17 +1189,17 @@ int UTIL_countPhysicalCores(void)
 
         free(buffer);
 
-        return numPhysicalCores;
+        return numCores;
     }
 
 failed:
     /* try to fall back on GetSystemInfo */
     {   SYSTEM_INFO sysinfo;
         GetSystemInfo(&sysinfo);
-        numPhysicalCores = sysinfo.dwNumberOfProcessors;
-        if (numPhysicalCores == 0) numPhysicalCores = 1; /* just in case */
+        numCores = sysinfo.dwNumberOfProcessors;
+        if (numCores == 0) numCores = 1; /* just in case */
     }
-    return numPhysicalCores;
+    return numCores;
 }
 
 #elif defined(__APPLE__)
@@ -1189,24 +1208,24 @@ failed:
 
 /* Use apple-provided syscall
  * see: man 3 sysctl */
-int UTIL_countPhysicalCores(void)
+int UTIL_countCores(int logical)
 {
-    static S32 numPhysicalCores = 0; /* apple specifies int32_t */
-    if (numPhysicalCores != 0) return numPhysicalCores;
+    static S32 numCores = 0; /* apple specifies int32_t */
+    if (numCores != 0) return numCores;
 
     {   size_t size = sizeof(S32);
-        int const ret = sysctlbyname("hw.physicalcpu", &numPhysicalCores, &size, NULL, 0);
+        int const ret = sysctlbyname(logical ? "hw.logicalcpu" : "hw.physicalcpu", &numCores, &size, NULL, 0);
         if (ret != 0) {
             if (errno == ENOENT) {
                 /* entry not present, fall back on 1 */
-                numPhysicalCores = 1;
+                numCores = 1;
             } else {
-                perror("zstd: can't get number of physical cpus");
+                perror("zstd: can't get number of cpus");
                 exit(1);
             }
         }
 
-        return numPhysicalCores;
+        return numCores;
     }
 }
 
@@ -1215,16 +1234,16 @@ int UTIL_countPhysicalCores(void)
 /* parse /proc/cpuinfo
  * siblings / cpu cores should give hyperthreading ratio
  * otherwise fall back on sysconf */
-int UTIL_countPhysicalCores(void)
+int UTIL_countCores(int logical)
 {
-    static int numPhysicalCores = 0;
+    static int numCores = 0;
 
-    if (numPhysicalCores != 0) return numPhysicalCores;
+    if (numCores != 0) return numCores;
 
-    numPhysicalCores = (int)sysconf(_SC_NPROCESSORS_ONLN);
-    if (numPhysicalCores == -1) {
+    numCores = (int)sysconf(_SC_NPROCESSORS_ONLN);
+    if (numCores == -1) {
         /* value not queryable, fall back on 1 */
-        return numPhysicalCores = 1;
+        return numCores = 1;
     }
 
     /* try to determine if there's hyperthreading */
@@ -1238,7 +1257,7 @@ int UTIL_countPhysicalCores(void)
 
         if (cpuinfo == NULL) {
             /* fall back on the sysconf value */
-            return numPhysicalCores;
+            return numCores;
         }
 
         /* assume the cpu cores/siblings values will be constant across all
@@ -1271,13 +1290,13 @@ int UTIL_countPhysicalCores(void)
             ratio = siblings / cpu_cores;
         }
 
-        if (ratio && numPhysicalCores > ratio) {
-            numPhysicalCores = numPhysicalCores / ratio;
+        if (ratio && numCores > ratio && !logical) {
+            numCores = numCores / ratio;
         }
 
 failed:
         fclose(cpuinfo);
-        return numPhysicalCores;
+        return numCores;
     }
 }
 
@@ -1288,52 +1307,70 @@ failed:
 
 /* Use physical core sysctl when available
  * see: man 4 smp, man 3 sysctl */
-int UTIL_countPhysicalCores(void)
+int UTIL_countCores(int logical)
 {
-    static int numPhysicalCores = 0; /* freebsd sysctl is native int sized */
-    if (numPhysicalCores != 0) return numPhysicalCores;
+    static int numCores = 0; /* freebsd sysctl is native int sized */
+#if __FreeBSD_version >= 1300008
+    static int perCore = 1;
+#endif
+    if (numCores != 0) return numCores;
 
 #if __FreeBSD_version >= 1300008
-    {   size_t size = sizeof(numPhysicalCores);
-        int ret = sysctlbyname("kern.smp.cores", &numPhysicalCores, &size, NULL, 0);
-        if (ret == 0) return numPhysicalCores;
+    {   size_t size = sizeof(numCores);
+        int ret = sysctlbyname("kern.smp.cores", &numCores, &size, NULL, 0);
+        if (ret == 0) {
+            if (logical) {
+                ret = sysctlbyname("kern.smp.threads_per_core", &perCore, &size, NULL, 0);
+                /* default to physical cores if logical cannot be read */
+                if (ret == 0)
+                    numCores *= perCore;
+            }
+
+            return numCores;
+        }
         if (errno != ENOENT) {
-            perror("zstd: can't get number of physical cpus");
+            perror("zstd: can't get number of cpus");
             exit(1);
         }
         /* sysctl not present, fall through to older sysconf method */
     }
+#else
+    /* suppress unused parameter warning */
+    (void) logical;
 #endif
 
-    numPhysicalCores = (int)sysconf(_SC_NPROCESSORS_ONLN);
-    if (numPhysicalCores == -1) {
+    numCores = (int)sysconf(_SC_NPROCESSORS_ONLN);
+    if (numCores == -1) {
         /* value not queryable, fall back on 1 */
-        numPhysicalCores = 1;
+        numCores = 1;
     }
-    return numPhysicalCores;
+    return numCores;
 }
 
 #elif defined(__NetBSD__) || defined(__OpenBSD__) || defined(__DragonFly__) || defined(__CYGWIN__)
 
 /* Use POSIX sysconf
  * see: man 3 sysconf */
-int UTIL_countPhysicalCores(void)
+int UTIL_countCores(int logical)
 {
-    static int numPhysicalCores = 0;
+    /* suppress unused parameter warning */
+    (void) logical;
 
-    if (numPhysicalCores != 0) return numPhysicalCores;
+    static int numCores = 0;
 
-    numPhysicalCores = (int)sysconf(_SC_NPROCESSORS_ONLN);
-    if (numPhysicalCores == -1) {
+    if (numCores != 0) return numCores;
+
+    numCores = (int)sysconf(_SC_NPROCESSORS_ONLN);
+    if (numCores == -1) {
         /* value not queryable, fall back on 1 */
-        return numPhysicalCores = 1;
+        return numCores = 1;
     }
-    return numPhysicalCores;
+    return numCores;
 }
 
 #else
 
-int UTIL_countPhysicalCores(void)
+int UTIL_countCores(int logical)
 {
     /* assume 1 */
     return 1;
@@ -1341,6 +1378,16 @@ int UTIL_countPhysicalCores(void)
 
 #endif
 
+int UTIL_countPhysicalCores(void)
+{
+    return UTIL_countCores(0);
+}
+
+int UTIL_countLogicalCores(void)
+{
+    return UTIL_countCores(1);
+}
+
 #if defined (__cplusplus)
 }
 #endif

diff --git a/programs/util.h b/programs/util.h
index f146f3a08aa5547b9e3450aa2e9c6335e0617865..c1c9634343b226be144ab01c7e0fd4e83b2eb47f 100644 (file)
--- a/programs/util.h
+++ b/programs/util.h
@@ -289,13 +289,19 @@ void UTIL_refFilename(FileNamesTable* fnt, const char* filename);
 FileNamesTable*
 UTIL_createExpandedFNT(const char* const* filenames, size_t nbFilenames, int followLinks);
 
+#if defined(_WIN32) || defined(WIN32)
+DWORD CountSetBits(ULONG_PTR bitMask);
+#endif
 
 /*-****************************************
  *  System
  ******************************************/
 
+int UTIL_countCores(int logical);
+
 int UTIL_countPhysicalCores(void);
 
+int UTIL_countLogicalCores(void);
 
 #if defined (__cplusplus)
 }

diff --git a/programs/zstd.1.md b/programs/zstd.1.md
index ae50928667d9e8495151b5461883e850c13c927c..066bcb83cef0560de02b3f7626534e5d8bae5068 100644 (file)
--- a/programs/zstd.1.md
+++ b/programs/zstd.1.md
@@ -125,6 +125,9 @@ the last one takes effect.
     This mode is the only one available when multithread support is disabled.
     Single-thread mode features lower memory usage.
     Final compressed result is slightly different from `-T1`.
+* `--auto-threads={physical,logical} (default: physical)`:
+    When using a default amount of threads via `-T0`, choose the default based on the number
+    of detected physical or logical cores.
 * `--adapt[=min=#,max=#]` :
     `zstd` will dynamically adapt compression level to perceived I/O conditions.
     Compression level adaptation can be observed live by using command `-v`.

diff --git a/programs/zstdcli.c b/programs/zstdcli.c
index 5dba729a147ab22145be63d618f769e351c58839..e496852f3fbb1a7732a386624a9f9cb62be0ea12 100644 (file)
--- a/programs/zstdcli.c
+++ b/programs/zstdcli.c
@@ -213,6 +213,7 @@ static void usage_advanced(const char* programName)
     DISPLAYOUT( " -T#    : spawns # compression threads (default: 1, 0==# cores) \n");
     DISPLAYOUT( " -B#    : select size of each job (default: 0==automatic) \n");
     DISPLAYOUT( "--single-thread : use a single thread for both I/O and compression (result slightly different than -T1) \n");
+    DISPLAYOUT( "--auto-threads={physical,logical} (default: physical} : use either physical cores or logical cores as default when specifying -T0 \n");
     DISPLAYOUT( "--rsyncable : compress using a rsync-friendly method (-B sets block size) \n");
 # endif
     DISPLAYOUT( "--exclude-compressed: only compress files that are not already compressed \n");
@@ -761,6 +762,9 @@ int main(int const argCount, const char* argv[])
         separateFiles = 0,
         setRealTimePrio = 0,
         singleThread = 0,
+#ifdef ZSTD_MULTITHREAD
+        defaultLogicalCores = 0,
+#endif
         showDefaultCParams = 0,
         ultra=0,
         contentSize=1;
@@ -950,6 +954,15 @@ int main(int const argCount, const char* argv[])
                 if (longCommandWArg(&argument, "--target-compressed-block-size=")) { targetCBlockSize = readSizeTFromChar(&argument); continue; }
                 if (longCommandWArg(&argument, "--size-hint=")) { srcSizeHint = readSizeTFromChar(&argument); continue; }
                 if (longCommandWArg(&argument, "--output-dir-flat")) { NEXT_FIELD(outDirName); continue; }
+#ifdef ZSTD_MULTITHREAD
+                if (longCommandWArg(&argument, "--auto-threads")) {
+                    const char* threadDefault = NULL;
+                    NEXT_FIELD(threadDefault);
+                    if (strcmp(threadDefault, "logical") == 0)
+                        defaultLogicalCores = 1;
+                    continue;
+                }
+#endif
 #ifdef UTIL_HAS_MIRRORFILELIST
                 if (longCommandWArg(&argument, "--output-dir-mirror")) { NEXT_FIELD(outMirroredDirName); continue; }
 #endif
@@ -1156,8 +1169,13 @@ int main(int const argCount, const char* argv[])
 #ifdef ZSTD_MULTITHREAD
     if ((nbWorkers==0) && (!singleThread)) {
         /* automatically set # workers based on # of reported cpus */
-        nbWorkers = UTIL_countPhysicalCores();
-        DISPLAYLEVEL(3, "Note: %d physical core(s) detected \n", nbWorkers);
+        if (defaultLogicalCores) {
+            nbWorkers = UTIL_countLogicalCores();
+            DISPLAYLEVEL(3, "Note: %d logical core(s) detected \n", nbWorkers);
+        } else {
+            nbWorkers = UTIL_countPhysicalCores();
+            DISPLAYLEVEL(3, "Note: %d physical core(s) detected \n", nbWorkers);
+        }
     }
 #else
     (void)singleThread; (void)nbWorkers;

diff --git a/tests/playTests.sh b/tests/playTests.sh
index e3a8ebd20352f7bd755f7a20a0b12d4e0d2f577b..6d202c715587d12c3ff573bf93b4ab5d26ccc38e 100755 (executable)
--- a/tests/playTests.sh
+++ b/tests/playTests.sh
@@ -1351,6 +1351,8 @@ if [ -n "$hasMT" ]
 then
     println "\n===>  zstdmt round-trip tests "
     roundTripTest -g4M "1 -T0"
+    roundTripTest -g4M "1 -T0 --auto-threads=physical"
+    roundTripTest -g4M "1 -T0 --auto-threads=logical"
     roundTripTest -g8M "3 -T2"
     roundTripTest -g8M "19 --long"
     roundTripTest -g8000K "2 --threads=2"