Tag line tag line Perforce Benchmark with PAM

Tag line, tag line Perforce Benchmark with PAM over NFS, FCP & i. SCSI Bikash R. Choudhury

Perforce Testing with Performance Acceleration Module ¡Goal – Compare Perforce benchmark results with and without PAM – Compare Perforce benchmark results over NFS v/s i. SCSI v/s FCP ¡Measured performance of metadata on Net. App storage with and without PAM – Entire feature set not considered – Tests conducted under controlled environment © 2008 Net. App. All rights reserved. 2

Executive Summary ¡ Result – 40% to 90% READ performance improvements with PAM – For READ performance, NFS slightly faster than i. SCSI and FCP without PAM ¡ Application is single threaded; i. SCSi/FCP on ext 3 with 4 k block size is less effective than NFS with 64 k blocks ¡ Host side cache gets flushed; NFS (file access) is faster than i. SCSI (blocks) served out of WAFL cache – For WRITE performance, i. SCSI is 7 x faster than NFS – Typical Perforce workload : 70% READ, 30% WRITE © 2008 Net. App. All rights reserved. 3

Deltas Benchmark – READ workload INTEGRATE Test 0. 2 88% Time in Seconds 0. 4 60 2 0. 6 0. 4 91% 0. 2 96% 0 Time in Seconds 0. 4 0 40 20 76% 36% 0. 8 Time in Seconds 75% 2. 5 0. 6 FAS 3070 With PAM 1. 2 80 0 FAS 3070 without PAM 1. 6 1 0. 2 FCP 60 50 40 30 20 10 0 Time in Seconds i. SCSI 0. 6 Time in Seconds 0. 8 1. 5 43% 1 0. 5 0 0 60 50 40 30 20 10 1. 6 1. 2 85% Time in Seconds NFS SYNC Test Time in Seconds DIR Test 0. 8 61% 0. 4 0 Response time improvements are averages across 2, 6, and 3 tests for Dir, Integrate and Sync test respectively © 2008 Net. App. All rights reserved. 4

Branchsubmit Benchmark – WRITE workload ¡ The test does "submit" of the change list action (p 4 submit -c <changenumber>) that holds approximately a dozen locks on the database tables for update purposes. – FCP ¡ 11 seconds – Elapsed time (time taken for each command to complete) ¡ 13500 files/sec – Commit rate (rate @ which files are written to disk) – i. SCSI ¡ 14 seconds – Elapsed time ¡ 14, 000 files/sec – Commit rate – NFS ¡ 50 seconds – Elapsed Time ¡ 2000 files/sec – Commit rate © 2008 Net. App. All rights reserved. 5

Perforce Architecture © 2008 Net. App. All rights reserved. 6

Perforce Architecture ¡Metadata DB – Files that compose the Perforce database ¡Journal Files - A record of all transactions performed by the server after the last checkpoint was created ¡Versioned Files (Depot) - Store the file revisions. ¡LOG – Contain the error output files © 2008 Net. App. All rights reserved. 7

Test Descriptions ¡ Deltas Benchmark (READ test) – Used Net. App Eng. IT real production data to generate the p 4 database and journal files for this test – DIR, INTEGRATE, SYNC tests generate READ workload ¡ INTEGRATE / DIR scans/parses/reads the different revisions of the code from the p 4 database with and without use of locks ¡ SYNC test downloads selected file(s) from the DEPOT to their client workspace – Odd and Even-numbered Integrates use related data ¡ Related data already in Net. App cache (memory + PAM) for evennumbered integrates – Demonstrates great performance improvement when data is already present in Net. App cache ¡ Branchsubmit Benchmark (WRITE Test) – Used Perforce’s Reference Dataset – Heavy WRITEs with limited READs; Single threaded ¡ Sets exclusive locks on files for writes ¡ Faster writes release locks for user read requests © 2008 Net. App. All rights reserved. 8

PAM Settings ¡There are three modes of operations for PAM – Default – Metadata – Low Priority ¡Mainly used in RANDOM read workloads ¡The actual application data is seldom reused in a timely manner – hence caching helps!! – The random nature of the p 4 workload tends to use more metadata © 2008 Net. App. All rights reserved.

Net. App Storage Specifications Net. App Storage Controller FAS 3070 – AA Cluster OS Data. ONTAP Release 7. 3 Cache 8 GB Net. App Drive Shelves One shelf with fourteen 300 Gb 15 k drives on each head Drive Access Protocol ISCSI NFS – Both over 1 gigabit interface FCP PAM 1 PCI Slot (16 GB) © 2008 Net. App. All rights reserved. 10

Perforce Host Specifications OS Redhat Enterprise Linux 5 update 2 Kernel 2. 6. 18 -88. el 5 xen #1 SMP NFS mount options rw, nfsvers=3, bg, hard, rsize=65536, wsize=65536, acregmin=3600, acregmax=3600, acdirmin=7200, acdirmax=7200, proto=tcp, nointr, nolock, timeo=600, retrans=5 Initiator for i. SCSI Software initiator Memory 32 Gb Processor(s) 8 way Intel Xeon quad Core 2. 00 GHz © 2008 Net. App. All rights reserved. 11

Hardware Connections ¡ Two filesystems mounted over NFSv 3 from one FAS 3070 head with 1 PAM ¡ Two i. SCSI LUNs mounted from another FAS 3070 head with 1 PAM card ¡ Two FCP LUNs mounted from same FAS 3070 head with 1 PAM card ¡ Both NFS & i. SCSI connections are over single dedicated 1 Gb Ethernet ¡ The FCP target and initiator were connected back-toback ¡ The NFS filesystems, i. SCSI and FCP LUNs were mounted to one single host running Perforce 2007. 3 on RHEL 5 update 2. © 2008 Net. App. All rights reserved. 12

Key Takeaways - PAM ¡PAM is needed where you want to reach high performance – – With few spindles in workloads With small blocks Random read workload A lot of metadata READ ops © 2008 Net. App. All rights reserved.

Thank You © 2008 Net. App. All rights reserved. 14
- Slides: 14