Re: performance boundaries

[RB <aoz.syn@xxxxxxxxx>]

On Fri, Nov 21, 2008 at 05:26, Huszár Viktor Dénes <hvd@xxxxxx> wrote:
> falls down to 800Mbit/s and the loop0 device uses only once cpu on 100%
<snip>
> is there any patch, trick, translation option which can brake the
> 800Mbit/sec borderline?

Seems you've already identified the bottleneck: you're CPU-bound.
Faster processors, more of them, or a crypto offload engine are about
your only way of increasing that speed.  Since you've gotten this far
I presume you already have, but  if you're using an off-the-shelf
Linux distro and x86_64 processors make sure your kernel & loop-aes
module properly match the hardware.

You also seem to have already done this, but for public edification:
make sure you use 'lvcreate -i N' when creating your volume like this,
where N is the number of loopback devices you have.  Otherwise, your
data will be allocated linearly and still limited to single-processor
speed.  It still will be for accesses up to 512k unless you changed
--stripesize.

Probably nothing new; CPU-bound is CPU-bound.  Unless you switch to a
"cheaper" cipher or buy more hardware, 800Mbps per core is likely
where you'll remain.


RB
âžì\¯*m¡Êò¦Ú ­ªaÊ)ڝÚ'¶‹Š{±³+-z`+r¯zm§ÿæj)g–X§»+ƒùbžì\¯*m£