On Wed, Jan 29, 2014 at 2:16 AM, Mel Gorman <mgorman@xxxxxxx> wrote: > On Tue, Jan 28, 2014 at 09:30:25AM -0800, Andy Lutomirski wrote: >> On Thu, Jan 16, 2014 at 4:56 PM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote: >> > I'm interested in a persistent memory track. There seems to be plenty >> > of other emails about this, but here's my take: >> >> I should add that I'm also interested in topics relating to the >> performance of mm and page cache under various abusive workloads. >> These include database-like things and large amounts of locked memory. >> > > Out of curiousity, is there any data available on this against a recent > kernel? Locked memory should not cause the kernel to go to hell as the > pages should end up on the unevictable LRU list. If that is not happening, > it could be a bug. More details on the database configuration and test > case would also be welcome as it would help establish if the problem is > a large amount of memory being dirtied and then an fsync destroying the > world or something else. > On (IIRC) 3.5, this stuff worked very poorly. On 3.9, with a lot of extra memory in the system, I seem to do okay. I'm planning on trying 3.13 with a more moderate amount of memory soon. On 3.11, with normal amounts of memory, something is still not so good. I'm seeing this on development boxes, so it may be filesystem-dependent. The performance things I actually care about lately are more in the category of getting decent page-fault performance on locked pages. Even better would be no page faults at all, but that may be a large project. The database in question is a proprietary thing (which I hope to open-source some day) that creates and fallocates 10-20MB files, mlocks them, reads a byte from every page to prefault them, then writes them once, reads them quite a few times, and, after a while, reads them one last time and deletes them. At any given time, there are a couple GB of these files open and mlocked. Performance seems to be okay (modulo page faults) once everything is up and running, but, at startup, the system can go out to lunch for a while. On a completely different workload (Mozilla Thunderbird's "Compact Now" button on btrfs), something certainly still destroys the world. I suspect that complaining about pathological cases in btrfs will get me nowhere, though... :-/. --Andy > -- > Mel Gorman > SUSE Labs -- Andy Lutomirski AMA Capital Management, LLC -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
