 |
 |
We are doing a lot of work on 3Ware tuning at RAL in order to optimise performance for the CCRC. Our main area of concern has been the terrible performance in concurrent read/write. Basically all the I/O seems to go to the writers. Few things make a difference, but the blockdev tweak up does improve the read performance and we have also seen good results with the anticipatory scheduler (at the expense of overall throughput). Bad concurrent read is a real killer for us as it hits migration to tape. I expect James Jackson (cc:) who is doing the work will blog/wiki a lot of our tests shortly. Also interested in other peoples experience. Regards Andrew > -----Original Message----- > From: GRIDPP2: Deployment and support of SRM and local storage > management [mailto:[log in to unmask]]On Behalf Of > John Bland > Sent: 29 February 2008 15:06 > To: [log in to unmask] > Subject: Re: 3ware Tuning > > > Andrew Elwell wrote: > > We've just got a ~48TB storage box here at Glasgow. - > 2*3ware cards with > > 24*1T disks on each. > > > > Painfully slow out of the box (it may be verifying just now - tw_cli > > unresponsive) but I discovered this snippet on the 3ware website > > > > http://www.3ware.com/kb/article.aspx?id=11050 > > > > gives some tuing tips. Anyone else have some to share? > > I did some benchmarking on a 16drive 3ware setup: > > http://map2.ph.liv.ac.uk/2007/04/13/raid5-performance-tweaking/ > > Mostly covers the 3ware tips and tuning xfs. > > The performance gains are geared towards small numbers of large files, > lots of small files will want the simultaneous requests and read-ahead > reduced. > > Out of the box the array should be fine unless it's > initialising in the > background, but tw_cli should still be mostly responsive no > matter what. > > If you use the 3DM2 web interface (might be in tw_cli as well > but I can > never remember where, and default password is 3ware) in Controller > configuration for each unit you can specify a performance level > Performance, Balanced or Protection, which boils down to how paranoid > you are about power loss during writes and whether you have a battery > backup unit. Performance does make a big difference on writes but only > recommended if you have a BBU on the card and UPS on the server. > > You can also specify the rate of initialisation/rebuild and verifying > compared to normal io operations. I tend to set the rebuild as highest > priority (want that sucker redundant again asap) and verifies as low > priority (an extra hour or two is unlikely to change whether the array > is broken or not right now). > > John > > -- > Dr John Bland, Systems Administrator > Room 210, Oliver Lodge > Particle Physics Group, University of Liverpool > Mail: [log in to unmask] > Tel : 0151 794 3396 >