Varnish memory usage

[Guillaume Quintard]

Hello Luca,

This should give you some information:
https://info.varnish-software.com/blog/understanding-varnish-cache-memory-usage
. Notably, you need to account for some memory fragmentation.

Note that your formula is probably a bit optimistic: each thread is allowed
to consume as much as workspace_client/workspace_backend, and workspace
will be consumed by VCL operations, like regex matching and substitutions.

One thing you can do is use vmod_vtc:
https://varnish-cache.org/docs/trunk/reference/vmod_vtc.html#int-workspace-free-enum-client-backend-session-thread
. You could log the remaining size for all requests and see how close you
get to zero.

I'd like to ask: which Varnish packages are you using? (i.e. version and
origin). I know some people have been hurt before, by the jemalloc package
I think, and depending on how Varnish is built, you can get better a
smaller memory footprint.

-- 
Guillaume Quintard


On Mon, Jul 1, 2024 at 5:02 PM Luca Gervasi <luca.gervasi at gmail.com> wrote:

> Hello,
> I'm following an issue that caused some OOMs during a huge spike.
> Currently I have the following properties set:
>
>  -p thread_pool_max=2000 \
> -p thread_pool_min=35 \
> -p feature=+esi_disable_xml_check \
> -p feature=+esi_ignore_other_elements \
> -p feature=+esi_remove_bom \
> -p feature=+http2 \
> -p pcre_match_limit_recursion=64 \
> -p syslog_cli_traffic=off \
> -p sigsegv_handler=on \
> -p workspace_client=2m \
> -p workspace_backend=2m \
> -p http_max_hdr=128 \
> -p http_req_hdr_len=32k \
> -p http_resp_hdr_len=32k \
> -p http_req_size=2m \
> -p http_resp_size=2m \
> -p thread_pool_stack=256k \
> -s malloc,20G \
> -S /config/varnish/secret \
> -f /config/varnish/default.vcl \
> -p vsl_reclen=700
>
> And I'm using this formula to somehow calculate how much memory is varnish
> going to use:
> thread_pool_max*thread_pool_stack+thread_pool_max*(workspace_client+http_req_hdr_len+http_req_size) +
> malloc
>
> And it is somewhere close to where I get OOM. I suspect that what did
> screw up was the high esi-to-request ration that I normally have:
>
> MAIN.client_req        469.22
> MAIN.esi_req          5489.25
>
> Hit rate on that esi requests is pretty high as we use high ttls and
> invalidations, but I think that those ESIs have to be taken into account
> too (as well as the backend requests too...).
>
> I wasn't able to find an official memory sizing guide, sorry. How do you
> size correctly memory usage?
> I know I should use MSE, at the moment is not on my list.
>
> Thanks
> Luca
> _______________________________________________
> varnish-misc mailing list
> varnish-misc at varnish-cache.org
> https://www.varnish-cache.org/lists/mailman/listinfo/varnish-misc
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://www.varnish-cache.org/lists/pipermail/varnish-misc/attachments/20240701/79568804/attachment-0001.html>