Re: [PATCH v1 2/2] zram: remove init_lock in zram_make_request

[Minchan Kim]

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 from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/