On 3/04/19 11:00 am, David Touzeau wrote: > > Le 02/04/2019 à 18:06, Alex Rousskov a écrit : >> On 4/2/19 1:23 AM, David Touzeau wrote: >>> Le 01/04/2019 à 23:22, Alex Rousskov a écrit : >>>> Do your Squids use shared memory for the memory cache? See >>>> memory_cache_shared (even if you do not set it explicitly). >>>> http://www.squid-cache.org/Doc/config/memory_cache_shared/ >>> The test did not use workers >> That does not answer my question. Do you use Rock cache_dir(s)? >> >> >>>> Any significant difference in mgr:info and mgr:counters output after a >>>> test that only has memory hits? >> The question still stands. I would recommend testing this with a single >> URL and a fixed/same number of requests submitted by a reliable proxy >> benchmarking tool or at least a wget/curl script. >> >> >>> Do you know why CentOS objects are 34 bytes smaller than Debian ? >> Something in your test setup or environment results responses (or >> response delivery statistics) that differ in size between the tests. I >> do not know what it is, and the number of possible options is too large >> to guess correctly: It could be anything from 32-vs-64 bit OSes, to >> locale differences, to Squid host name, to Cookies, to test setup >> imperfections, to Squid statistics collection bugs, etc., etc. >> >> Have you compared the responses received by the client (headers and >> all)? Do they differ by 34 bytes? I suggest testing with a single URL >> that produces different results and then digging down to identify the >> difference (starting with comparing responses). >> >> Alex. > > Thanks Alex for these ways to investigate. > > We will try to get more precise for the tests > > We have reduced the squid.conf to the minimal way in order to be sure > that nothing can disturb testing. > > Only one cache, no rock, no workers, no tuning > > Amos says that perhaps the C++ version can make some tweaks in this case > we will start to do the same tests with Ubuntu that uses the most recent > kernels and C++ > > But for the moment > > - without intelligence > - Without investigations efforts > - Just by compiling squid > > To resume > > Centos 7 is 10 times faster than Debian 7 > Centos 7 is 400% faster than Debian 9 > > Debian 9 is a little faster than Debian 7 > Can you provide the same numbers for Deb 7 that you did for Deb 9 ? so we can see what exactly "a little faster" means. PS. your first post which said "10 times faster" also said it was for "Debian 9 net install" - no Debian 7. > We will keep updated for Unbuntu. > Thanks. Amos _______________________________________________ squid-users mailing list squid-users@xxxxxxxxxxxxxxxxxxxxx http://lists.squid-cache.org/listinfo/squid-users