Hello, We have an application that does a huge amount of disk i/o (mostly writes) and have hit performance problems. It seems that the use of O_DIRECT would help but information on the development status is hard to come by. I have checked 2.4.18 and 2.5 kernels and there seems to be an interface but no actual implementation. It has been suggested that it only works with ext2 but we need ext3. Could someone tell me what the actual position with it is? Thanks. -- Ian Leonard eMail: ileonard@ntlworld.com Phone: +44 (0)1865 765273 Please ignore spelling and punctuation - I did.
Hi, On Thu, May 09, 2002 at 11:55:32AM +0100, Ian Leonard wrote:> I have checked 2.4.18 and 2.5 kernels and there seems to > be an interface but no actual implementation. > > It has been suggested that it only works with ext2 but we > need ext3.That's right. But the ext2 is pretty much only a wrapper around a generic_direct_IO function, so it would be pretty easy to use that same wrapper for ext3. It would be tricky to make it work with data=journal mode, but the other data modes ought to be easy to get going. Cheers, Stephen
On 2002.05.10 21:31 Stephen C. Tweedie wrote:> Hi, > > On Thu, May 09, 2002 at 11:55:32AM +0100, Ian Leonard wrote: > > > I have checked 2.4.18 and 2.5 kernels and there seems to > > be an interface but no actual implementation. > > > > It has been suggested that it only works with ext2 but we > > need ext3. > > That's right. > > But the ext2 is pretty much only a wrapper around a generic_direct_IO > function, so it would be pretty easy to use that same wrapper for > ext3. It would be tricky to make it work with data=journal mode, but > the other data modes ought to be easy to get going.I my case this wouldn't matter too much. I am trying to improve the performance of an application that writes video data to the disk. I have mounted the relevant filesystem with various data= options but it doesn't seem to make any difference to my test program (which simply writes a large file to disk and times it). I am far from familiar with the internals of filing systems. Do you think this feature is likely to appear in the near future. Also would it be a 2.4 or a 2.5 thing? Currently through-put doesn't look too bad but I see that there is a drop off every 30 seconds. I assumed that this was bdflush and fiddled with the options but haven't seen an improvement yet. Thanks for your help. Any words of wisdom appreciated. -- Ian Leonard eMail: ileonard@ntlworld.com Phone: +44 (0)1865 765273 Please ignore spelling and punctuation - I did.