Merge pull request #2925 from embg/dict_training_sample_limit_size

Allow user to specify memory limit for dictionary training

programs/dibio.c

@@ -309,7 +309,7 @@ static fileStats DiB_fileStats(const char** fileNamesTable, int nbFiles, size_t
 int DiB_trainFromFiles(const char* dictFileName, size_t maxDictSize,
                        const char** fileNamesTable, int nbFiles, size_t chunkSize,
                        ZDICT_legacy_params_t* params, ZDICT_cover_params_t* coverParams,
-                       ZDICT_fastCover_params_t* fastCoverParams, int optimize)
+                       ZDICT_fastCover_params_t* fastCoverParams, int optimize, unsigned memLimit)
 {
     fileStats fs;
     size_t* sampleSizes; /* vector of sample sizes. Each sample can be up to SAMPLESIZE_MAX */
@@ -341,6 +341,11 @@ int DiB_trainFromFiles(const char* dictFileName, size_t maxDictSize,
         /* Limit the size of the training data to 2GB */
         /* TODO: there is opportunity to stop DiB_fileStats() early when the data limit is reached */
         loadedSize = (size_t)MIN( MIN((S64)maxMem, fs.totalSizeToLoad), MAX_SAMPLES_SIZE );
+        if (memLimit != 0) {
+            DISPLAYLEVEL(2, "!  Warning : setting manual memory limit for dictionary training data at %u MB \n",
+                (unsigned)(memLimit / (1 MB)));
+            loadedSize = (size_t)MIN(loadedSize, memLimit);
+        }
         srcBuffer = malloc(loadedSize+NOISELENGTH);
         sampleSizes = (size_t*)malloc(fs.nbSamples * sizeof(size_t));
     }

programs/dibio.h

@@ -34,6 +34,6 @@
 int DiB_trainFromFiles(const char* dictFileName, size_t maxDictSize,
                        const char** fileNamesTable, int nbFiles, size_t chunkSize,
                        ZDICT_legacy_params_t* params, ZDICT_cover_params_t* coverParams,
-                       ZDICT_fastCover_params_t* fastCoverParams, int optimize);
+                       ZDICT_fastCover_params_t* fastCoverParams, int optimize, unsigned memLimit);
 
 #endif

programs/zstd.1.md

@@ -190,6 +190,10 @@ the last one takes effect.
 
     This is also used during compression when using with --patch-from=. In this case,
     this parameter overrides that maximum size allowed for a dictionary. (128 MB).
+
+    Additionally, this can be used to limit memory for dictionary training. This parameter
+    overrides the default limit of 2 GB. zstd will load training samples up to the memory limit
+    and ignore the rest.
 * `--stream-size=#` :
     Sets the pledged source size of input coming from a stream. This value must be exact, as it
     will be included in the produced frame header. Incorrect stream sizes will cause an error.
@@ -329,6 +333,8 @@ Compression of small files similar to the sample set will be greatly improved.
     resulting in a _small_ compression ratio improvement for this level.
 * `-B#`:
     Split input files into blocks of size # (default: no split)
+* `-M#`, `--memory=#`:
+    Limit the amount of sample data loaded for training (default: 2 GB). See above for details.
 * `--dictID=#`:
     A dictionary ID is a locally unique ID
     that a decoder can use to verify it is using the right dictionary.

programs/zstdcli.c

@@ -1327,18 +1327,18 @@ int main(int argCount, const char* argv[])
             int const optimize = !coverParams.k || !coverParams.d;
             coverParams.nbThreads = (unsigned)nbWorkers;
             coverParams.zParams = zParams;
-            operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenames->fileNames, (int)filenames->tableSize, blockSize, NULL, &coverParams, NULL, optimize);
+            operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenames->fileNames, (int)filenames->tableSize, blockSize, NULL, &coverParams, NULL, optimize, memLimit);
         } else if (dict == fastCover) {
             int const optimize = !fastCoverParams.k || !fastCoverParams.d;
             fastCoverParams.nbThreads = (unsigned)nbWorkers;
             fastCoverParams.zParams = zParams;
-            operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenames->fileNames, (int)filenames->tableSize, blockSize, NULL, NULL, &fastCoverParams, optimize);
+            operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenames->fileNames, (int)filenames->tableSize, blockSize, NULL, NULL, &fastCoverParams, optimize, memLimit);
         } else {
             ZDICT_legacy_params_t dictParams;
             memset(&dictParams, 0, sizeof(dictParams));
             dictParams.selectivityLevel = dictSelect;
             dictParams.zParams = zParams;
-            operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenames->fileNames, (int)filenames->tableSize, blockSize, &dictParams, NULL, NULL, 0);
+            operationResult = DiB_trainFromFiles(outFileName, maxDictSize, filenames->fileNames, (int)filenames->tableSize, blockSize, &dictParams, NULL, NULL, 0, memLimit);
         }
 #else
         (void)dictCLevel; (void)dictSelect; (void)dictID;  (void)maxDictSize; /* not used when ZSTD_NODICT set */

tests/playTests.sh

@@ -1051,6 +1051,13 @@ then
 fi
 rm -f tmp* dictionary
 
+println "- Test --memory for dictionary compression"
+datagen -g12M -P90 > tmpCorpusHighCompress
+zstd --train -B2K tmpCorpusHighCompress -o tmpDictHighCompress --memory=10K && die "Dictionary training should fail : --memory too low (10K)"
+zstd --train -B2K tmpCorpusHighCompress -o tmpDictHighCompress --memory=5MB 2> zstTrainWithMemLimitStdErr
+cat zstTrainWithMemLimitStdErr | grep "setting manual memory limit for dictionary training data at 5 MB"
+cat zstTrainWithMemLimitStdErr | grep "Training samples set too large (12 MB); training on 5 MB only..."
+rm zstTrainWithMemLimitStdErr
 
 println "\n===>  fastCover dictionary builder : advanced options "
 TESTFILE="$PRGDIR"/zstdcli.c