ah; previous message was an example of 'during the operation' i.e., randread entire sets; sorry about that, cmds uses only a tiny MB ~100MB once files stored. for cosd, during the operation uses 150-200mb per cosd (both during write/read) still, when considering performance, previous message stands or else the performance decreases if the minimum amount of RAM isn't there. cheers On Sun, Jun 12, 2011 at 20:41, djlee064 <djlee064@xxxxxxxxx> wrote: > based on the empirical measurement, starting with a set of (real) > number of files, focusing mainly on small-files. and found some rough > relationship, > all filesets with same dist. i.e., ~80% <4MB file, mostly small, all 1x. replc. > For MDS RAM: > > You need about: > 0.6GB RAM to store 0.03million files (fileset vol. 1.2TB) > 1.2GB to store 0.065million files (fileset vol. 2.4TB) > 1.8GB to store 0.13million files (fileset vol. 4.8TB) > > ratio of no.files per GB fortunately increases, (i.e., 0.07million per > GB for the 0.13million files), hopefully per-GB will support > 0.1million files as more files are stored. > then, 18GB = 1.8million files, > 180GB = 18million > 1800GB= 180million (this is about 6.64PB), > > So to support that amount,, we need 100 MDS nodes with 18GB purely for > cmds, incl mem for OS, etc, maybe 20-24GB per node. > > Cheers > > On Sat, Jun 11, 2011 at 06:47, Fyodor Ustinov <ufm@xxxxxx> wrote: >> >> Hi! >> >> Which configuration would you recommended for cluster with 50-80 million files? >> >> WBR, >> Fyodor. >> -- >> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in >> the body of a message to majordomo@xxxxxxxxxxxxxxx >> More majordomo info at http://vger.kernel.org/majordomo-info.html > -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
