On 2016-12-23 13:07 (+0100), Daniel Shahaf <danie...@apache.org> wrote: > Jakub Petr wrote on Thu, Dec 22, 2016 at 10:28:24 +0100: > > Why the "transmitting phase" should take so long? > > Usually it's because of virus scanners intercepting the disk operations > (opening/writing files). > > What makes you think the slowness is svnserve's fault? Is the commit > faster over file:// than over svn://localhost?
There is more that can be tried to reduce I/O and to check whether there it results in improved performance: * use 'svn import' instead of add/commit (no pristine store update) * in fsfs.conf, disable rep-sharing (no temp. sha1 files) * 'svnadmin create --compatible-version 1.8' (no index file updates) * add '-c 0' to svnserve's command line (may save ~1 minute) * add '-M 100' to svnserve's command line (make sure large flat directories are cached properly) * distribute the files over multiple directories - in case they are currently in one "flat" directory -- Stefan^2.
