80 likes | 295 Views
4720 24 drives Seq. access test. Targeting for Video post production use Focusing on sequential access only (not the SSD advantage are) Compared host base 2 stripe LVM * 2 and 4 stripe LVM with XFS
E N D
472024 drives Seq. access test • Targeting for Video post production use • Focusing on sequential access only (not the SSD advantage are) • Compared host base 2 stripeLVM *2 and 4 stripeLVM with XFS • DH4720 can hold 24* SFF(2.5”)drives on 2U height chassis, having good space factor when performance requirements met • Created vdisks with RAID5,chunk size 512KB • Linux host(SuSE12.3, since some-customer was using) • 1M stripe size had performance advantage for RAID5*2 or 4LVM • # lvcreate -L capacity -i 2 or 4-I 1m-n lvnamevgname • DMMP configured 1pref+1unprefpaths, tests using 1 path/pv • 2M stripe was over-sized • XFS gives better throughput than WindowsNTFS with smaller I/O bock size • With not deep QD, like two, almost best performing(can be universally applied for many type of applications) Dot Hill Systems
Configuration1. 2d-LVM*2 3TB R5 3TB R5 DellR610 SuSELinux12.3 QLE2564 HBA IOmeter 2010 Create striped VG using two PVs from same controller (2* workers). Create200GBxfs with several stripe size. 3TB R5 3TB R5 4720 24 drives 1 chassis FC-DAS 8Gb*2 or 4 1 or 2 vd / port 600GB 10K SAS *6 600GB 10K SAS *6 FRUKA19 (shorty v2 chassis) KF64 (SFF 600GB 10K SAS) Hitachi HUC1060* x 24 600GB 10K SAS *6 600GB 10K SAS *6 Two * RAID5(A-side) Two * RAID5(B-side) 1*preferred path + 1*other path per volume (Using 1* preferred path only) 4720 (redundant) FRUKC50 4* 8G FC,Linear FX, 4GB Arrandale 1.8GHz SC Vertex-6 FPGA (Jerboa) Dot Hill Systems
2d-LVM*2xfsresults • These show 2 volumes total • (6dRAID5*2 stripe, 12 drives on LV) * 2, 24 drives total • One volume performance is half of these • 2d-LVM*2 in total was best. Considering one volume only, 4d-LVM win. • If AP can do IO with QD=2, 16KB block I/O gives good performance. Dot Hill Systems
Configuration2.4d-LVM 3TB R5 3TB R5 DellR610 SuSELinux 12.3 QLE2564 HBA IOmeter 2010 Created host based 4 stripe LVM. Createxfsand200GBIOBW.tstfiles for each VG. Stripe size 256K, 512K, and 1M. 3TB R5 3TB R5 4720 and 4824 24 drives 1 chassis FC-DAS 8Gb*2 or 4 1 or 2 vd / port 600GB 10K SAS *6 600GB 10K SAS *6 FRUKA19 (shorty v2 chassis) KF64 (SFF 600GB 10K SAS) Hitachi HUC1060* x 24 600GB 10K SAS *6 600GB 10K SAS *6 Two * RAID5(A-side) Two * RAID5(B-side) 4720 (redundant) FRUKC50 4* 8G FC,Linear FX, 4GB Arrandale 1.8GHz SC Vertex-6 FPGA (Jerboa) Dot Hill Systems
4d LVM xfs 性能比較 • Even with QD=1, >=64KB block IO can achieve >=1,100MB/s • Both for Read or Write access • With QD=2, >=16KB block IO can achieve >=1,150MB/s • Similar result with QD=4 or 8 Dot Hill Systems
2d-LVMvs.4d-LVM • Graph resized to be able to directly compare 1 volume performance • If best single stream performance needed, need to use all 24 drives, configure 4d-LVM.Use 1M byte stripe size to create LV on XFS. • With 12 drives 2d-LVM, not so poor. • Has advantage when allover goodness. Performance, usable capacity and more WS can be supported. Dot Hill Systems
2d-LVMvs.4d-LVM (2) • Two drives LVM *2 gives best total throughput • All are using 24 drives. RAID5(5d+1p)*4RAID sets. • Using 8Gb*4 connections. Logical B/W max ~= 3,200MB/s Dot Hill Systems
ref)Linux vs. Windows Linux (left) vs.Windows(right) • LinuxXFS seems using more aggressive cache • NTFSsector size: default used 8KB? Dot Hill Systems