single threaded testing will always limit your ability to get any higher throughput. Can you try multiple threads of say 10 from one client and add/remove as you see fit? How did you arrive at 30MB/s throughput requirement?? Reason I ask is that in most cases it's not well thought through and you try to test way way more than what's required. Do you get 5000 operations/sec, with 6k file and 30MB/s this is what you would see. If that's the case I don't think gluster will scale unless you add more nodes and use lots of clients to test. Where did you run "dd"? On gluster client mount? On Tue, May 17, 2011 at 7:24 PM, nuaa_liuben <nuaa_liuben at sina.com> wrote: > Avg. files size is between 2 ~ 6KB, both text and binary files. > 1) about 30 MB/s for singel client, this can be realized? > 2) dd with 4KB can get about 30 - 40MB/s > 3) multiple clients can?near linearly increase performance > 4) only use single thread for testing > > BR, > liuben > >>From:? Mohit Anchlia >>Send Time: ?2011-05-18? 09:54:45 >>TO: nuaa_liuben >>CC: ?gluster-users >>Subject: Re: Gluster 3.2 for lots of small files >> > You?will?find?difficult?to?tune?when?using?small?files.?But?what?you > are?seeing?is?really?low.?Is?your?average?file?size?10k??Are?these > text?files? > 1)?of?all?what's?your?throughput?requirements?? > 2)?try?running?"dd"?tests?with?oflag/iflag=direct > 3)?run?multiple?clients?from?separate?hosts?and?see?if?you?add?a > client?does?it?linearly?increase?the?performance. > 4)?when?using?small?files?increasing?number?of?parallel?threads?from > same?client?will?not?increase?performance.?lower?the?threads?and?see > if?that?helps. > Post?results. > On?Tue,?May?17,?2011?at?6:23?PM,?nuaa_liuben?<nuaa_liuben at sina.com>?wrote: >>?Hi, >> >>?I?am?testing?gluster?3.2?for?LOSF?(lots?of?small?files)?scenario?recently, >>?with?iozne?and?postmark?benchmark?tools.?And?the?results?show?that?LOSF >>?performance?is?really?bad,?random?read/write?rate?is?just?only?about?2?~ >>?4MB/s.?I?tried?RAID0,?RAID1,?RAID10,?and?it's?the?same?result.?And?create >>?ops?with?postmark?is?100?~?400?per?second. >> >>?I?deployed?4?bricks?servers:?Dell?R510,??4?1TB?SATA?disks,?1?GiE?NIC,?8GB >>?Mem,?and?test?machine?is?also?R510?with?2?SAS?disks?and?16GB?memory. >> >>?Any?good?ideas?on?gluster3.2?tunning?for?LOSF? >>?Why?dbd?translator?is?removed?from?gluster?3.2? >>?thanks?in?advance. >> >> >>?BR, >>?liuben >>?_______________________________________________ >>?Gluster-users?mailing?list >>?Gluster-users at gluster.org >>?http://gluster.org/cgi-bin/mailman/listinfo/gluster-users >> >>
