On Thu, Jan 29, 2015 at 03:06:04PM +0900, Sergey Senozhatsky wrote: > On (01/29/15 14:28), Minchan Kim wrote: > > > I'm still concerned about performance numbers that I see on my x86_64. > > > it's not always, but mostly slower. I'll give it another try (disable > > > lockdep, etc.), but if we lose 10% on average then, sorry, I'm not so > > > positive about srcu change and will tend to vote for your initial commit > > > that simply moved meta free() out of init_lock and left locking as is > > > (lockdep warning would have been helpful there, because otherwise it > > > just looked like we change code w/o any reason). > > > > > > what do you thunk? > > > > Surely I agreee with you. If it suffers from 10% performance regression, > > it's absolutely no go. > > > > However, I believe it should be no loss because that's one of the reason > > from RCU birth which should be really win in read-side lock path compared > > to other locking. > > > > Please test it with dd or something for block-based test for removing > > noise from FS. I also will test it to confirm that with real machine. > > > > do you test with a single dd thread/process? just dd if=foo of=bar -c... or > you start N `dd &' processes? I tested it with multiple dd processes. > > for a single writer there should be no difference, no doubt. I'm more > interested in multi-writer/multi-reader/mixed use cases. > > the options that I use are: iozone -t 3 -R -r 16K -s 60M -I +Z > and -I is: > -I Use VxFS VX_DIRECT, O_DIRECT,or O_DIRECTIO for all file operations > > with O_DIRECT I don't think there is a lot of noise, but I'll try to use > different benchmarks a bit later. > As you told, the data was not stable. Anyway, when I read down_read implementation, it's one atomic instruction. Hmm, it seems te be better for srcu_read_lock which does more things. But I guessed most of overhead are from [de]compression, memcpy, clear_page That's why I guessed we don't have measurable difference from that. What's the data pattern if you use iozone? I guess it's really simple pattern compressor can do fast. I used /dev/sda for dd write so more realistic data. Anyway, if we has 10% regression even if the data is simple, I never want to merge it. I will test it carefully and if it turns out lots regression, surely, I will not go with this and send the original patch again. Thanks. > > -ss -- Kind regards, Minchan Kim -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>