On 11/21/23 16:00, Andres Freund wrote:
> Hi,
>
> On 2023-11-21 14:48:59 -0400, David Steele wrote:
>>> I'd not call 7.06->4.77 or 6.76->4.77 "virtually free".
>>
>> OK, but how does that look with compression
>
> With compression it's obviously somewhat different - but that part is done in
> parallel, potentially on a different machine with client side compression,
> whereas I think right now the checksumming is single-threaded, on the server
> side.
Ah, yes, that's certainly a bottleneck.
> With parallel server side compression, it's still 20% slower with the default
> checksumming than none. With client side it's 15%.
Yeah, that still seems a lot. But to a large extent it sounds like a
limitation of the current implementation.
>> -- to a remote location?
>
> I think this one unfortunately makes checksums a bigger issue, not a smaller
> one. The network interaction piece is single-threaded, adding another
> significant use of CPU onto the same thread means that you are hit harder by
> using substantial amount of CPU for checksumming in the same thread.
>
> Once you go beyond the small instances, you have plenty network bandwidth in
> cloud environments. We top out well before the network on bigger instances.
>
>> Uncompressed backup to local storage doesn't seem very realistic. With gzip
>> compression we measure SHA1 checksums at about 5% of total CPU.
>
> IMO using gzip is basically infeasible for non-toy sized databases today. I
> think we're using our users a disservice by defaulting to it in a bunch of
> places. Even if another default exposes them to difficulty due to potentially
> using a different compiled binary with fewer supported compression methods -
> that's gona be very rare in practice.
Yeah, I don't use gzip anymore, but there are still some platforms that
do not provide zstd (at least not easily) and lz4 compresses less. One
thing people do seem to have is a lot of cores.
>> I can't understate how valuable checksums are in finding corruption,
>> especially in long-lived backups.
>
> I agree! But I think we need faster checksum algorithms or a faster
> implementation of the existing ones. And probably default to something faster
> once we have it.
We've been using xxHash to generate checksums for our block-level
incremental and it is seriously fast, written by the same guy who did
zstd and lz4.
Regards,
-David