On Wed, Sep 17, 2008 at 01:44:16AM +0200, Arno Wagner wrote: > hat do you want multicore support for ayways? Do you have an > application that can actually saturate even one core? Multicore > sounds like basicelly costing a lot of effort and postential > problems for nothing in return. Yes, easily. I find quite often that kcryptd takes 100% of one CPU, including in non-artificial setups when working with huge files. Perhaps it's the 3 disk raid capable of 200 MiB/s, I don't know, that's just what I've seen watching top. Of course might be that dm-crypt is broken in some other way that makes it sometimes take lots of CPU. The effect would probably be more noticeable on some older SMP hardware; I have a Core 2 Quad. As an artificial benchmark that should probably give an upper bound on the speedup in my case: ------------------------------------------------------------ # hdparm -t /dev/mapper/myvg-root /dev/mapper/myvg-root: Timing buffered disk reads: 634 MB in 3.01 seconds = 210.90 MB/sec # hdparm -t /dev/mapper/myvg-root_crypt /dev/mapper/myvg-root_crypt: Timing buffered disk reads: 258 MB in 3.01 seconds = 85.79 MB/sec ------------------------------------------------------------ I don't know about the internals, but I'd assume multiple threads could also reduce read latency by breaking up large requests. Definitely dm-crypt noticeably slows down disk access, so I guess the question is how much of the performance could be gained back by multiple threads. At the moment my computer has been up for 22 days and ps shows kcryptd has so far taken 14.5 hours of CPU time, and I'm sure it's not evenly distributed over the uptime; this is not a database server or a high load web server or anything like that. Hm, you said there's currently one thread per device. So would it theoretically help if I ran raid (or lvm) over dm-crypt, not the other way round? Perhaps I should try that, although it somehow seems backwards. Sami --------------------------------------------------------------------- dm-crypt mailing list - http://www.saout.de/misc/dm-crypt/ To unsubscribe, e-mail: dm-crypt-unsubscribe@xxxxxxxx For additional commands, e-mail: dm-crypt-help@xxxxxxxx
