Phil Schwan
2006-May-19 07:36 UTC
[Lustre-discuss] Stable performance from a one disk OST?
Magnus Karlsson wrote:> > I have been using Lustre now for nearly a year, and the problem I have > is kind of odd. For a number of reasons I desire to have a relatively > stable performance (throughput and latency) from the file system. In > all the releases since 0.6 I have noticed that I indeed get this when > the majority of the files are in the memory of the OST, but not when > they are on disk. Every 60 seconds kswapd suddenly swaps in 200-700 > pages and the throughput drops an order of a magnitude and the latency > shoots up an order of magnitude for a second or two. Do you know how I > could alleviate this behavior?Over the last two years we have seen countless strange behaviours from the Linux VM. I can''t say for sure whether we have seen this particular one. We do have one patch, which can probably be included in 1.0.3, which will let you turn off the OST''s read cache by flipping a bit in /proc. I would be curious to hear if this resolves your issue. Can you try again after 1.0.3 and get back to us? Thanks-- -Phil
Magnus Karlsson
2006-May-19 07:36 UTC
[Lustre-discuss] Stable performance from a one disk OST?
Hi, I have been using Lustre now for nearly a year, and the problem I have is kind of odd. For a number of reasons I desire to have a relatively stable performance (throughput and latency) from the file system. In all the releases since 0.6 I have noticed that I indeed get this when the majority of the files are in the memory of the OST, but not when they are on disk. Every 60 seconds kswapd suddenly swaps in 200-700 pages and the throughput drops an order of a magnitude and the latency shoots up an order of magnitude for a second or two. Do you know how I could alleviate this behavior? Here are my system details: Lustre 1.0.1 on your standard 2.4.20 kernel rpm. The OST is running on a box with 2 1GHz P3 processors and with 2G of memory. There is only one UltraSCSI2 HP disk drive connected to the system. The load I generate using iozone is of one thread sequentially reading a file (in 64K chunks) that is larger than 2G (so that it will be on disk). One possible solution I could use as a last resort is to buy one extra HD for every single box that I have, and then have the Lustre file system data on one, and RedHat on the other. I would be great if somebody could suggest something cheaper ;-). BTW, Lustre has come a long way in a year. The 1.0 release is really good. Congratulations. Thanks: Magnus Karlsson