Load-Sto-Meter: Generating Workloads for Persistent Memory Damini Chopra, Doug Voigt Hewlett Packard (Enterprise)

Transcription

1 Load-Sto-Meter: Generating Workloads for Persistent Memory Damini Chopra, Doug Voigt Hewlett Packard (Enterprise)

2 Application vs. Pure Workloads Benchmarks that reproduce application workloads Assist in system provisioning Assist in tuning for nuances of particular applications Get more accurate results Pure workload generators Assist in characterizing basic system performance Assist in locating root causes of performance issues Evaluate design trade-offs in prototype systems Explore best/worst application performance Enable finer grained analysis than profiling 2

3 Difference Between IO and PM Workloads Application User space Native file API Load/ store Kernel space NVM.PM.FILE mode NVM.PM.VOLUME mode PM-aware kernel module PM-aware file system MMU Mappings NVM PM capable driver PM device PM device... PM device Byte Addressability No software involved in CPU Load/Store instructions Additional sync operation Avoid context switch from user space to kernel space (no I/O syscalls) Eliminate expensive sync operations to disk by introducing CPU Memory syncs 3

4 Role of Pure Workload Generator Analyze the performance and behavior of Persistent Memory Tune the parameters of Load-Sto-Meter Simulate real application behavior Provide aggressive Multithreaded Benchmarking Analyze PM sync mechanism versus DISK sync in terms of performance and reliability over various parameterized workloads Evaluate PM sync implementations 4

5 NVM Programming Model Map and Sync User Addr User App Multithreads/processor Process Virtual Addr Map Linux Scheduler PM_SYNC DRAM PM DISK_SYNC FS I/O Scheduler Driver mmap DISK 5

6 Basic Ld/St/Sync Template No mem-access while SYNC mem-access Performance metric measured at every interval 6

7 Workload Generation Parameters Store to Sync ratio Read to Write Ratio Workload Threads Shared v/s Private memory access Sequential v/s Random memory access Granularity of Memory Access (Load/Store record size) PM Data Structures NUMA switch 7

8 Typical Output PM vs. Disk Sync Latency Latency (ns) File Size in Kby (Log) PM_SYNC DISK_SYNC Performance Metric Giga-Updates Per Second Latency Potential Measurement Domain Filer Size Granularity Thread Count Sync Ratio Latency (vs GUPS) 8

9 Abstracting Syncs Sync implementation depends on many things Disk vs. PM User vs. Kernel space Processor instruction set Additional features such as High Availability Performance is sensitive to NVM.PM.SYNC implementations NVM Programming Model specifies sync/optimized flush abstraction Application independent Processor architecture independent Implementation independent Workload generator should support multiple sync implementations User plug in feature for their own implementation of sync Easily test multiple PM aware file systems with the same workload 9

10 Multi-Threading Socket 0 Socket 1 Core 0 Workload Thread #0 Core 1 Workload Thread #1 Core 2 Workload Thread #2 L3 Cache Core 4 Measuring Thread Core 5 Core 6 Core 3 Core 7 Workload Thread #3 L2 Cache L2 Cache L2 Cache L2 Cache DISK_SYNC Persistent Memory DISK Periodic_sleep PM_SYNC Memory Load/Store Cached Data 10

11 Sources of Performance Noise Trade-off between performance and thermal management policy of OS Partial utilization of CPU, causing discrepancies in acquiring pure load/store performance Hyper-threading management, while trying to embark equal workload on each core of the processor 11

12 A new type of workload generator is needed! Performance measurement specification? Workload generation parameters Definition of parameter driven behavior Implementations? Commercial opportunity Open source opportunity SNIA SW TWG? Open source contributions? 12

13 Load-Sto-Meter: Generating Workloads for Persistent Memory Thank You