Hi. Decided to run the same benchmark on our CephFS filesystem backed by 72 x 7,2K RPM 10TB disks on 6 OSD-hosts connected through 10GbitE.
# /root/bench.sh With fadvise blocksize 4096 Issuing fadvise real 0m1.351s user 0m0.035s sys 0m0.236s Without fadvise blocksize 4096 real 0m8.604s user 0m0.055s sys 0m0.197s Ok, x6 speedup on 4KB blocks speed without: 10000*4096/8.6 => 4.5MB/s Speed with: 10000*4096/1.35 => 29.3 MB/s ... With fadvise blocksize 131072 Issuing fadvise real 0m1.580s user 0m0.222s sys 0m0.805s Without fadvise blocksize 131072 real 0m30.549s user 0m0.255s sys 0m1.040s Ok, speedup x19 Speed without: 131072*10000/30.549 => 40.8 MB/s Speed with : 131072*10000/1.580 => 791 MB/s Note that above tests are with page-cache flushed on the client side, but the data are still resident on the cache in the OSD-end. That one is hard to flush, but benefits should be even better on cold-cold storage - which bacula by its nature normally trigger. Tried to put some IO on the backend .. pushing the files out of local cache on the OSD-nodes (cannot be confirmed). Fadvise for 10K x 128K files - 7s Without for 10K x 128K files - 172s Jesper _______________________________________________ Bacula-users mailing list Bacula-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/bacula-users
