On 17 August 2012 09:36, Kent Overstreet [off-list ref] wrote:

On Thu, Aug 16, 2012 at 10:07:29PM +1000, Joseph Glanville wrote:

quoted

There also seems to have been pretty severe performance regressions in
cache bypassed sequential I/O.
The newer code barely does 70mb/s sequential writes when
sequential_cuttoff is set to 4M however it does around 300mb/s when
set to 0 (no bypass) with dd and 1M block size.
When using the older codebase the cache bypass is actually slightly
faster than going to cache at around 360mb/s but this is still much
slower than the underlying block devices (as previously discussed).

Ouch. Can you try profiling it with perf while running dd?

perf record -afg dd etc. etc.
perf report

should do it

Sure can.

-- 
CTO | Orion Virtualisation Solutions | www.orionvm.com.au
Phone: 1300 56 99 52 | Mobile: 0428 754 846