[Lustre-discuss] crc32 is only checksum avail?

Dilger, Andreas andreas.dilger at intel.com
Thu Jan 31 18:19:11 PST 2013


On 2013/28/01 3:22 PM, "Andrus, Brian Contractor" <bdandrus at nps.edu> wrote:
>I am using the lustre-2.2.0-2.6.32_220.4.2.el6_lustre.x86_64.x86_64 rpms
>downloaded from whamcloud.
>
>I checked the checksums on our system and find:
>lctl get_param osc.*.checksum_type
>osc.WORK-OST0000-osc-MDT0000.checksum_type=[crc32]
>osc.WORK-OST0001-osc-MDT0000.checksum_type=[crc32]
>osc.WORK-OST0002-osc-MDT0000.checksum_type=[crc32]
>osc.WORK-OST0003-osc-MDT0000.checksum_type=[crc32]
>osc.WORK-OST0004-osc-MDT0000.checksum_type=[crc32]
>osc.WORK-OST0005-osc-MDT0000.checksum_type=[crc32]
>osc.WORK-OST0006-osc-MDT0000.checksum_type=[crc32]
>osc.WORK-OST0007-osc-MDT0000.checksum_type=[crc32]
>osc.WORK-OST0008-osc-MDT0000.checksum_type=[crc32]
>osc.WORK-OST0009-osc-MDT0000.checksum_type=[crc32]
>osc.WORK-OST000a-osc-MDT0000.checksum_type=[crc32]
>osc.WORK-OST000b-osc-MDT0000.checksum_type=[crc32]
>osc.WORK-OST000c-osc-MDT0000.checksum_type=[crc32]
>osc.WORK-OST000d-osc-MDT0000.checksum_type=[crc32]
>
>Ok, so I understand crc32 is not the best, so I want to change it to
>adler.
>#echo adler > /proc/fs/lustre/osc/WORK-OST0000-osc-MDT0000/checksum_type
>-bash: echo: write error: Invalid argument
>
>Hmm? Is there something more to this? I would think the packages provided
>would have both checksum types available.

Hmm, strange.  Adler should always be available since the introduction of
checksums in 1.8.  It does also depend algorithm the server is supporting.
 Are you also using 2.2 servers, and were they compiled with Adler
support?  That depends on whether the kernel config if it is available.

In newer Lustre versions with proper CPU hardware support, CRC32 can
actually use less CPU than Adler, but I don't think this is the case for
2.2 clients/servers, only in 2.3 and later.

Cheers, Andreas
-- 
Andreas Dilger

Lustre Software Architect
Intel High Performance Data Division





More information about the lustre-discuss mailing list