[ovs-discuss] ovs-vswitchd memory consumption behavior
blp at ovn.org
Wed Mar 20 18:22:28 UTC 2019
OK. Thanks for the update.
asan and valgrind are very general tools. There are other tools, or at
least there used to be, which sometimes have fewer requirements, for
helping with leaks specifically. I have not used any of them in the
last few years because valgrind and asan have worked so well for me, so
I don't have any of them to specifically recommend.
On Tue, Mar 19, 2019 at 11:56:42PM +0000, Fernando Casas Schössow wrote:
> The guys from musl confirmed ASan will not work with musl there are known issues.
> [00:13] <@dalias> the asan runtime doesn't work with musl afaik. it's very invasive and i'm not sure what would take to make it work
> [00:14] <nsz> yeah it does not work
> [00:18] <@dalias> it could just be the issue with musl's malloc wrongly enlarging the heap under heavy contention
> [00:18] <nsz> musl supports malloc interposition nowadays, so you can try another allocator maybe
> [00:18] <@dalias> *nod*
> [00:19] <nsz> try ldpreloading jemalloc or tcmalloc and see if the problem disappears
> Problem is jemalloc nor tcmalloc are available on Alpine out of the box.
> So I will check if I can compile jemalloc and test OVS using ld preloading as suggested.
> On mar, mar 19, 2019 at 7:18 PM, Fernando Casas Schössow <casasfernando at outlook.com> wrote:
> I discussed this with one of Alpine developers and it's not clear that ASan works with musl.
> Also opened an issue to Alpine about this to see if I can get also some help from their side:
> I will check in musl developers IRC channel to try to confirm if using ASan is an option or I will be wasting my time. :(
> On mar, mar 19, 2019 at 8:13 AM, Fernando Casas Schössow <casasfernando at outlook.com> wrote:
> Not yet. I will look into how to do it in Alpine.
> On lun, mar 18, 2019 at 10:37 PM, Ben Pfaff <blp at ovn.org> wrote:
> Did you try compiling with Address Sanitizer yet? It is just as effective as valgrind at finding leaks, but it slows the process much less than valgrind does, so it is less likely to change behavior due to timing issues.
More information about the discuss