[SPDK] Re: iSCSI Sequential Read with BS 128K Performance

Hi Lego,

If you want to test, but not changing the code, you may try  the following patch:

https://review.spdk.io/gerrit/c/spdk/spdk/+/5664 It supports to leverage RPC to configure the small and large buffer pool size.




Best Regards,
Ziye Yang

-----Original Message-----
From: Yang, Ziye 
Sent: Tuesday, December 22, 2020 1:46 PM
To: Storage Performance Development Kit <spdk(a)lists.01.org>
Subject: RE: [SPDK] iSCSI Sequential Read with BS 128K Performance

Hi Lego,

" Focus on BS 128K
[OK] QD1: read: IOPS=5543, BW=693MiB/s (727MB/s) [OK] QD8: read: IOPS=21.1k, BW=2634MiB/s (2762MB/s) [NOK] QD16: read: IOPS=301, BW=37.7MiB/s (39.5MB/s) "

How about changing 

#define BUF_LARGE_POOL_SIZE			1023   in lib/bdev/bdev.c

This Macro may have impact on the performance if you testing big I/O size test. We currently define 1023 * 64 KB with the total size.  So Would you change the value to a bigger value, and see whether it helps the test? // Currently, we did not export it to configurable yet.


Best Regards,
Ziye Yang

-----Original Message-----
From: Lego Lin <lego.lin(a)gmail.com>
Sent: Friday, December 18, 2020 4:09 PM
To: Storage Performance Development Kit <spdk(a)lists.01.org>
Subject: [SPDK] iSCSI Sequential Read with BS 128K Performance

Hi, all:
I just tested SPDK 20.10.x iSCSI performance and I got following performance data: (All Testing with FIO QD 32)
[OK] (100%RndR_4K,   100%RndW_4K)   : (read: IOPS=155k, BW=605MiB/s
(634MB/s),   write: IOPS=140k, BW=547MiB/s (573MB/s))
[OK] (100%RndR_8K,   100%RndW_8K)   : (read: IOPS=147k, BW=1152MiB/s
(1208MB/s), write: IOPS=128k, BW=1003MiB/s (1051MB/s)) [NOK] (100%SeqR_128K, 100%SeqW_128K) : (read: IOPS=210, BW=26.3MiB/s (27.5MB/s),  write: IOPS=14.6k, BW=1831MiB/s (1920MB/s)) => Read Bad, Write OK [NOK] (100%SeqR_32K,  100%SeqR_16K)  : (read: IOPS=9418, BW=294MiB/s
(309MB/s),   read: IOPS=105k, BW=1641MiB/s (1721MB/s))
=> [NOK] BS_32K + QD_32
=> [OK]  BS_16K + QD_32
[OK] (100%SeqR_8K,   100%SeqR_4K)   : (read: IOPS=149k, BW=1160MiB/s
(1217MB/s), read: IOPS=157k, BW=612MiB/s (642MB/s))

Focus on BS 128K
[OK] QD1: read: IOPS=5543, BW=693MiB/s (727MB/s) [OK] QD8: read: IOPS=21.1k, BW=2634MiB/s (2762MB/s) [NOK] QD16: read: IOPS=301, BW=37.7MiB/s (39.5MB/s)

FIO Configuration:
ioengine=libaio
direct=1
numjobs=1

I also check with document:
https://ci.spdk.io/download/performance-reports/SPDK_tcp_perf_report_2007.pdf
Inside this document, it also suggest Sequential Read test with BS=128K +
QD8

I think the low performance with BS=128K + QD32 should not relate to iSCSI, but can anyone share experience about tuning iSCSI sequential read performance? It's weird that performance drop with high QD. Any suggestion are welcome.

Thanks

Following are my test configuration
1. Network bandwidth: 40GB
2. TCP Setting at both target and client
    tcp_timestamps: "1"
    tcp_sack: "0",
    tcp_rmem: "4096 87380 134217728"
    tcp_wmem: "4096 87380 134217728",
    tcp_mem: "4096 87380 134217728",
    rmem_default: "524287",
    wmem_default: "524287",
    rmem_max: "268435456",
    wmem_max: "268435456",
    optmem_max: "268435456",
    netdev_max_backlog: "300000"}
3. number of CPU cores at target and client: 48 vcores
     Intel(R) Xeon(R) Silver 4215 CPU @ 2.50GHz * 2 4. disable irqbalance / enable CPU power government 5. run SPDK with: ./iscsi_tgt -m 0x08007C08007C _______________________________________________
SPDK mailing list -- spdk(a)lists.01.org
To unsubscribe send an email to spdk-leave(a)lists.01.org