s3: support random fill in object prefix

A sequence of three of more "%" chars in an object prefix will now get
replaced by a random hex string of the same length.

CHANGELOG.md

@@ -1,5 +1,14 @@
 # Changelog of elbencho
 
+## v3.0.6 (work in progress)
+
+### New Features & Enancements
+* Added new multi-arch (ARM64 & x86_64) docker container with CUDA/GDS support.
+* If an S3 object prefix contains a sequence of three or more '%' chars, this sequence will now get replaced by a random hex string of the same length.
+
+### Contributors
+* Thanks to Phil Canman, Erez Binia and Ohad Shamir for helpful comments and suggestions.
+
 ## v3.0.5 (Jan 05, 2024)
 
 ### New Features & Enhancements

Makefile

@@ -5,7 +5,7 @@
 EXE_NAME           ?= elbencho
 EXE_VER_MAJOR      ?= 3
 EXE_VER_MINOR      ?= 0
-EXE_VER_PATCHLEVEL ?= 5
+EXE_VER_PATCHLEVEL ?= 6
 EXE_VERSION        ?= $(EXE_VER_MAJOR).$(EXE_VER_MINOR)-$(EXE_VER_PATCHLEVEL)
 EXE                ?= $(BIN_PATH)/$(EXE_NAME)
 EXE_UNSTRIPPED     ?= $(EXE)-unstripped

source/ProgArgs.cpp

@@ -475,7 +475,8 @@ void ProgArgs::defineAllowedArgs()
 			"Don't check for S3 multi-part uploads exceeding 10,000 parts.")
 /*s3o*/	(ARG_S3OBJECTPREFIX_LONG, bpo::value(&this->s3ObjectPrefix),
 			"S3 object prefix. This will be prepended to all object names when the benchmark path "
-			"is a bucket.")
+			"is a bucket. (A sequence of 3 to 16 \"" RAND_PREFIX_MARKS_SUBSTR "\" chars will be "
+			"replaced by a random hex string of the same length.)")
 /*s3r*/	(ARG_S3RANDOBJ_LONG, bpo::bool_switch(&this->useS3RandObjSelect),
 			"Read at random offsets and randomly select a new object for each S3 block read. Only "
 			"effective in read phase and in combination with \"-" ARG_NUMDIRS_SHORT "\" & \"-"
@@ -685,6 +686,7 @@ void ProgArgs::defineDefaults()
 	this->madviseFlags = 0;
 	this->runS3MultiDelObjNum = 0;
 	this->disablePathBracketsExpansion = false;
+	this->useS3ObjectPrefixRand = false;
 }
 
 /**
@@ -748,6 +750,7 @@ void ProgArgs::initImplicitValues()
 		blockVariancePercent = 0;
 	}
 
+	useS3ObjectPrefixRand = (s3ObjectPrefix.find(RAND_PREFIX_MARKS_SUBSTR) != std::string::npos);
 }
 
 /**
@@ -2958,6 +2961,9 @@ void ProgArgs::setFromPropertyTreeForService(bpt::ptree& tree)
 		gpuIDsStr = gpuIDsServiceOverride;
 
 	parseGPUIDs();
+
+	// init implict vals
+	useS3ObjectPrefixRand = (s3ObjectPrefix.find(RAND_PREFIX_MARKS_SUBSTR) != std::string::npos);
 }
 
 /**

source/ProgArgs.h

@@ -204,6 +204,10 @@ namespace bpt = boost::property_tree;
 #define ARG_MADVISE_FLAG_NOHUGEPAGE_NAME	"nohugepage"
 
 
+#define RAND_PREFIX_MARK_CHAR				'%' // name prefix char to replace with random value
+#define RAND_PREFIX_MARKS_SUBSTR			"%%%" // three times RAND_PREFIX_MARK_CHAR
+
+
 typedef std::vector<CuFileHandleData> CuFileHandleDataVec;
 typedef std::vector<CuFileHandleData*> CuFileHandleDataPtrVec;
 
@@ -372,6 +376,7 @@ class ProgArgs
 		unsigned short s3LogLevel; // log level for AWS SDK
 		bool noDirectIOCheck; // ignore directIO alignment and sanity checks
 		std::string s3ObjectPrefix; // object name/path prefix for s3 "directory mode"
+		bool useS3ObjectPrefixRand; // implicit based on RAND_PREFIX_MARKS_SUBSTR in s3ObjectPrefix
 		uint64_t runS3ListObjNum; // run seq list objects phase if >0, given number is listing limit
 		bool runS3ListObjParallel; // multi-threaded object listing (requires "-n" / "-N")
 		bool doS3ListObjVerify; // verify object listing (requires "-n" / "-N")
@@ -559,7 +564,8 @@ class ProgArgs
 		bool getUseS3TransferManager() const { return useS3TransferManager; }
 		unsigned short getS3LogLevel() const { return s3LogLevel; }
 		bool getNoDirectIOCheck() const { return noDirectIOCheck; }
-		std::string getS3ObjectPrefix() const { return s3ObjectPrefix; }
+		const std::string& getS3ObjectPrefix() const { return s3ObjectPrefix; }
+		bool getUseS3ObjectPrefixRand() const { return useS3ObjectPrefixRand; }
 		uint64_t getS3ListObjNum() const { return runS3ListObjNum; }
 		bool getRunListObjPhase() const { return (runS3ListObjNum > 0); }
 		bool getRunListObjParallelPhase() const { return runS3ListObjParallel; }

source/toolkits/random/RandAlgoGoldenPrime.h

@@ -25,6 +25,11 @@ class RandAlgoGoldenPrime : public RandAlgoInterface
 			state = stateSeeder.next();
 		}
 
+		RandAlgoGoldenPrime(uint64_t seed)
+		{
+			state = seed;
+		}
+
 		virtual ~RandAlgoGoldenPrime() {}
 
 	private:

source/workers/LocalWorker.cpp

@@ -50,6 +50,8 @@
 #define NETBENCH_CONNECT_TIMEOUT_SEC	20 // max time for servers to wait and clients to retry
 #define NETBENCH_RECEIVE_TIMEOUT_SEC	20 // max time to wait for incoming data on client & server
 #define NETBENCH_SHORT_POLL_TIMEOUT_SEC	2  // time to check for interrupts in longer poll wait loops
+#define HEX_ALPHABET					"0123456789ABCDEF"
+#define HEX_ALPHABET_LEN				(sizeof(HEX_ALPHABET) - 1)
 
 
 #ifdef S3_SUPPORT
@@ -3367,6 +3369,7 @@ void LocalWorker::s3ModeIterateObjects()
 		all workers use the dirs of worker rank 0 */
 	const bool useTransMan = progArgs->getUseS3TransferManager();
 	std::string objectPrefix = progArgs->getS3ObjectPrefix();
+	const bool objectPrefixRand = progArgs->getUseS3ObjectPrefixRand();
 	const BenchPhase globalBenchPhase = workersSharedData->currentBenchPhase;
 	const size_t localWorkerRank = workerRank - progArgs->getRankOffset();
 	const bool isRWMixedReader = ( (globalBenchPhase == BenchPhase_CREATEFILES) &&
@@ -3405,9 +3408,14 @@ void LocalWorker::s3ModeIterateObjects()
 					"dirIndex: " + std::to_string(dirIndex) + "; "
 					"fileIndex: " + std::to_string(fileIndex) );
 
+			if(objectPrefixRand)
+				objectPrefix = getS3RandObjectPrefix(
+					workerRank, dirIndex, fileIndex, progArgs->getS3ObjectPrefix() );
+
 			unsigned bucketIndex = (workerRank + dirIndex) % bucketVec.size();
 			std::string currentObjectPath = objectPrefix + currentPath.data();
 
+
 			rwOffsetGen->reset(); // reset for next file
 
 			std::chrono::steady_clock::time_point ioStartT = std::chrono::steady_clock::now();
@@ -3484,6 +3492,7 @@ void LocalWorker::s3ModeIterateObjectsRand()
 	const size_t workerDirRank = progArgs->getDoDirSharing() ? 0 : workerRank; /* for dir sharing,
 		all workers use the dirs of worker rank 0 */
 	std::string objectPrefix = progArgs->getS3ObjectPrefix();
+	const bool objectPrefixRand = progArgs->getUseS3ObjectPrefixRand();
 
 	// init random generators for dir & file index selection
 
@@ -3534,6 +3543,10 @@ void LocalWorker::s3ModeIterateObjectsRand()
 				"dirIndex: " + std::to_string(dirIndex) + "; "
 				"fileIndex: " + std::to_string(fileIndex) );
 
+		if(objectPrefixRand)
+			objectPrefix = getS3RandObjectPrefix(
+				workerRank, dirIndex, fileIndex, progArgs->getS3ObjectPrefix() );
+
 		const unsigned bucketIndex = (workerRank + dirIndex) % bucketVec.size();
 		std::string currentObjectPath = objectPrefix + currentPath.data();
 
@@ -4535,6 +4548,7 @@ void LocalWorker::s3ModeListObjParallel()
 	const size_t workerDirRank = progArgs->getDoDirSharing() ? 0 : workerRank; /* for dir sharing,
 		all workers use the dirs of worker rank 0 */
 	std::string objectPrefix = progArgs->getS3ObjectPrefix();
+	const bool objectPrefixRand = progArgs->getUseS3ObjectPrefixRand();
 	const bool doListObjVerify = progArgs->getDoListObjVerify();
 
 
@@ -4586,6 +4600,10 @@ void LocalWorker::s3ModeListObjParallel()
 					"dirIndex: " + std::to_string(dirIndex) + "; "
 					"fileIndex: " + std::to_string(fileIndex) );
 
+			if(objectPrefixRand)
+				objectPrefix = getS3RandObjectPrefix(
+					workerRank, dirIndex, fileIndex, progArgs->getS3ObjectPrefix() );
+
 			std::string currentObjectPath = objectPrefix + currentPath.data();
 
 			expectedObjs.insert(currentObjectPath);
@@ -4711,7 +4729,7 @@ void LocalWorker::s3ModeVerifyListing(StringSet& expectedSet, StringList& receiv
 }
 
 /**
- * List objects and add multi-delete them in given buckets with user-defined limit for number of
+ * List objects and multi-delete them in given buckets with user-defined limit for number of
  * entries.
  *
  * @throw WorkerException on error.
@@ -4840,6 +4858,45 @@ bool LocalWorker::getS3ModeDoReverseSeqFallback()
 	return false;
 }
 
+/**
+ * In S3 mode, replace any sequence of at least 3 consecutive RAND_PREFIX_MARK_CHAR chars with a
+ * random uppercase hex string based on worker rank, dir index and file index. It's based on these
+ * so that we can calculate the same random values again later to find the files.
+ *
+ * Note: It's a good idea to check progArgs->getUseS3ObjectPrefixRand() to avoid calling this
+ * 		unnecessairly.
+ *
+ * @objectPrefix string in which to repace the consecutive occurences of RAND_PREFIX_MARK_CHAR.
+ * @return objectPrefix with replaced RAND_PREFIX_MARK_CHAR chars.
+ */
+std::string LocalWorker::getS3RandObjectPrefix(size_t workerRank, size_t dirIdx,
+	size_t fileIdx, const std::string& objectPrefix)
+{
+	size_t threeMarksPos = objectPrefix.find(RAND_PREFIX_MARKS_SUBSTR);
+
+	if(threeMarksPos == std::string::npos)
+		return objectPrefix; // not found, so nothing to replace here
+
+	std::string randObjectPrefix(objectPrefix); // the copy to replace chars
+
+	// we don't want any zero-based to turn result to all-zero (e.g. "-n 0" would always be 0)
+	workerRank++;
+	dirIdx++;
+	fileIdx++;
+
+	uint64_t randomNum = RandAlgoGoldenPrime(workerRank * dirIdx * fileIdx).next();
+
+	for(size_t i = threeMarksPos;
+		(i < objectPrefix.size() ) && (objectPrefix[i] == RAND_PREFIX_MARK_CHAR);
+		i++)
+	{
+		randObjectPrefix[i] = ( (char*)HEX_ALPHABET)[randomNum % HEX_ALPHABET_LEN];
+
+		randomNum /= HEX_ALPHABET_LEN;
+	}
+
+	return randObjectPrefix;
+}
 
 /**
  * Return appropriate file open flags for the current benchmark phase in dir mode.

source/workers/LocalWorker.h

@@ -220,6 +220,8 @@ class LocalWorker : public Worker
 			std::string bucketName, std::string listPrefix);
 		void s3ModeListAndMultiDeleteObjects();
 		bool getS3ModeDoReverseSeqFallback();
+		std::string getS3RandObjectPrefix(size_t workerRank, size_t dirIdx, size_t fileIdx,
+			const std::string& objectPrefix);
 
 		void hdfsDirModeIterateDirs();
 		void hdfsDirModeIterateFiles();