Hi Mikhail, On Mon, Aug 23, 2021 at 05:57:15PM +0200, Mikhail Sennikovsky wrote: > For bulk ct entry loads (with -R option) reusing the same nftc handle > for all entries results in ~two-time reduction of entries creation > time. This becomes signifficant when loading tens of thouthand of > entries. This is showing on of the limitations of the original API, I started sketching a patch to update this code to use libmnl, I'd rather follow this path. Would you have the time to take it over and look into it? > E.g. in the tests performed with the tests/conntrack/bulk-load-stress.sh > the time used for loading of 131070 ct entries (2 * 0xffff) > was 1.05s when this single nfct handle adjustment and 1.88s w/o it . What is making things go faster? What is introduding the extra overhead? Thanks
