[LU-13391] no print checksum speed for t10pi checksum in debug messages Created: 26/Mar/20  Updated: 26/Mar/20

Status: Open
Project: Lustre
Component/s: None
Affects Version/s: Lustre 2.14.0
Fix Version/s: None

Type: Bug Priority: Minor
Reporter: Shuichi Ihara Assignee: WC Triage
Resolution: Unresolved Votes: 0
Labels: None
Environment:

master


Issue Links:
Related
is related to LU-13393 t10crc4K/512 algorithm in rhel8.1 ker... Resolved
Severity: 3
Rank (Obsolete): 9223372036854775807

 Description   

no print checksum speed for t10pi (e.g. t10ip512, t10ip4k, t10crc512, t10crc4k) in debug messages. It would print speed as well as crc32c, adler algorithms.

00000400:02000000:4.0F:1585184902.841590:0:6755:0:(libcfs_cpu.c:1234:cfs_cpu_init()) HW NUMA nodes: 1, HW CPU cores: 32, npartitions: 8
00000001:01000000:4.0:1585184903.092471:0:6755:0:(linux-crypto.c:390:cfs_crypto_performance_test()) Crypto hash algorithm adler32 speed = 2715 MB/s
00000001:01000000:4.0:1585184903.342242:0:6755:0:(linux-crypto.c:390:cfs_crypto_performance_test()) Crypto hash algorithm crc32 speed = 13779 MB/s
00000001:01000000:4.0:1585184903.592260:0:6755:0:(linux-crypto.c:390:cfs_crypto_performance_test()) Crypto hash algorithm crc32c speed = 16427 MB/s
00000020:02000000:20.0F:1585184932.491506:0:6767:0:(class_obd.c:627:obdclass_init()) Lustre: Build Version: 2.12.3_ddn23
00000800:02000000:20.0:1585184932.526790:0:6767:0:(o2iblnd.c:2587:kiblnd_hdev_get_attr()) Using FastReg for registration
00000400:02000000:20.0:1585184932.538991:0:6767:0:(api-ni.c:2225:lnet_startup_lndni()) Added LNI 10.0.11.1@o2ib10 [8/512/0/180]
00000800:02000000:20.0:1585184932.539125:0:6767:0:(o2iblnd.c:2587:kiblnd_hdev_get_attr()) Using FastReg for registration
00000400:02000000:20.0:1585184932.550103:0:6767:0:(api-ni.c:2225:lnet_startup_lndni()) Added LNI 10.20.11.1@o2ib20 [8/512/0/180]
 


 Comments   
Comment by Li Xi [ 26/Mar/20 ]

The reason for this is, T10pi checksums are not normal checksum types like crc32c or alder. So the performance calculations are done in a different (but similar and comparable) way in a different place. The performance will only be calculated and printed when being used for the first time. It would be nice to print the performance together with other algorithms.

Another more cirtical thing is, if T10PI checksum performance is much slower than other algorithms, a warning message should be printed. That is especially important for Lustre client which might be forced to use T10PI checksum even the performance is bad. An warning message would at least let the administrator notice the performance penalty.

Generated at Sat Feb 10 03:00:52 UTC 2024 using Jira 9.4.14#940014-sha1:734e6822bbf0d45eff9af51f82432957f73aa32c.