[LU-1256] Question about checksums

• Status: Resolved
• Project: Lustre
• Affects Version/s: Lustre 2.1.0
• Type: Improvement
• Priority: Minor
• Reporter: Roger Spellman (Inactive)
• Assignee: Cliff White (Inactive)
• Resolution: Fixed

Description

I have a few questions about Checksums.

Section 19.5.1 of the manual says:

To check the status of a wire checksum, run:

lctl get_param osc.*.checksums

Does that mean that clearing this parameter only clears the wire checksum?

Does writing each file named /proc/fs/lustre/osc/*/checksums do the same thing?

Is there a way to permanently disable checksums?

Thanks.

Comments

[Peter Jones]

Cliff

Could you please help out with this one?

Thanks

Peter

[Cliff White (Inactive)]

Are you using the current Lustre manual? I am checking the 2.0 manual, per 19.5.1.
Does seem to be a bit unclear, perhaps should be rewritten.

First, in general, the old way of parameter access is directly to the /proc files, the new
modern way is using get/set commands, if the path is the same both (atm) should do the same thing, but the get/set method
is future-compatible. You should be able to set a persistent value with lctl conf_param.

Both types of checksums are controlled by "lctl set_param llite.*.checksum_pages" which is a client-only parameter.

Network (wire) checksums can be set/unset with the osc.*.checksums parameter. This is done per-client and per ost connection.
Again, you can set/get with lctl or by direct file access.

So, you can have network and memory checksums, or memory only, or none.

[Roger Spellman (Inactive)]

Thanks, Cliff.

Is there a way to permanently disable checksums?

There is CRC on the wire, and ECC on the RAM.
Exactly what errors are the checksums supposed to catch?

[Cliff White (Inactive)]

As I said, you should be able to turn them off persistently with lctl conf_param, there is no other switch. CRC and ECC are hardware checksums, the Lustre checksums include the Lustre stack and are supposed to catch any issues in that area.

[Peter Jones]

As per Peter Piela this ticket can be closed