>Since you have written a tool to test performance, you can compare
>my aoeserver, kvblade (try both the raw bd and via loopback) and vblade.
>I think that you will come to the conclusion that vblade, aoeserver and
>kvblade+loopback are all be able to max out your ethernet/diskio. But
>thats just a guess :)
>> On the copying to/from kernel to user space, again I did some
>> performance analysis and found that while the fastest I can do disk IO
>> at 35MB/s, 'dparm -T' says that I can copy to/from kernel buffers at
>> around 1750MB/s which suggests that the memory copies aren't going to
>> have a significant effect on performance.
Not significant perhaps but certainly an effect. Don't want to state the
obvious but it's always better to let the CPU do real work instead of
just moving chunks of memory around. =3D)
Also, having a slow processor and several IO-cards and disks would change
that ratio quite a lot, probably enough to make the performance penalty
clearly noticable if you run a DBMS or webserver or whatever on the same
machine.. Perhaps not the most common case but still.
>After having written a kernel-based version I think the best way to
>go is a userland implementation :)
Agree, I'd like to port your aoeserver module to OpenBSD so.. Less work
for me. ;)
Trust is a complex social phenomenon
-- Bruce Schneier
PGP Fingerprint: CBD4 89F2 0935 BB81 63FA 4384 7A87 A930 D98E 2F4B)