filebench介绍.pdf

Richard McDougall Sun Microsystems FileBench A Prototype Model Based Workload for File Systems Work In Progress Copyright Sun Microsystems, 2004 2 Benchmarks? For Vendors Product characterization Product design goaling Benchmarketing For Customers Purchasing Guide Configuration characterization/tuning/verification Copyright Sun Microsystems, 2004 3 Requirements for file-level benchmarking Represent Apps rather than I/Os Trace-derived synthesis Thread-level representation Inter-thread dependency/sync. Forward Path Extensible to new protocols Modular to include test of client: process/thread model, cpu efficiency etc... Pre-structuring/aging of file sets Scalable Throughput, #Users #Files/Directories Working set size #Clients Client resources (mem/cpu) Copyright Sun Microsystems, 2004 4 NFS 3/4 POSIX CIFS? XML? OLTP ERP DSS Mail What do we want to characterize? File System St or ag e Se rv er Applications Access Method File System Client Storage Server Copyright Sun Microsystems, 2004 5 Characterization Strategies I/O Microbenchmarking Pros: Easy to run Cons: Small test coverage, Hard to correlate to real apps Trace Capture/Replay I/O Trace, NFS Trace, Application Trace Pros: Accurate reconstruction of real application I/O mix Cons: Large traces, difficult to reconstruct I/O dependencies Model Based Distillation of trace into representive model Probability based, Simulation based Pros: Easy to run, Scalable in multiple dimensions Cons: Care required to ensure accurate real-world representation Copyright Sun Microsystems, 2004 6 Model based methodology study Application Level Trace ●Thread ●File/Dir ●Attrs etc... Workload Model Workload Replay ●Scale Factors Measurement Target ●FS, Client, Server etc ●Measurement Attrs Copyright Sun Microsystems, 2004 7 Model Allows Complex/Important Scaling Curves e.g. Thoughput/Latency vs. Working set size Thoughput/Latency vs. #users CPU Efficiency vs. Thoughput Caching efficiency vs. Workingset size/Memsize Copyright