Terrible hme throughput
Frank Jahnke
jahnke at sonatabio.com
Wed Sep 27 17:43:40 PDT 2006
> fj> There's one or two factors of two left to be found. Maybe it
> fj> is the Sparc disadvantage for these sorts of calculations
>
> no, I don't think there are any more factors of two to find.
>
> 300MHz Pentium, Linux with gcc: 1.5MByte/s
> 440MHz UltraSPARC II, Solaris with Sun C compiler: 2.3MByte/s
> 500MHz UltraSPARC II, FreeBSD with gcc: 1.0MByte/s
>
> try a slow PeeCee and see if you get similar results. I think it's
> about right: divide performance in half as penalty for trying to use
> gcc on anything but i386.
I had no idea that there was so much overhead with scp. I tried scp to
localhost on my main workstation (dual Athlons, 15K SCSIs, FreeBSD) and
got about 7MB/s. That is a hit of about a factor of 10 for using scp
from the native disk rate.
I will indeed try it on a slow PC soon: I am bring up a dual Pentium III
server in a week so so.
>
> My friend who makes big ftp servers with dm_crypt encrypted disks
> reports results roughly in the same ballpark: 40MByte/s throughput
> IDE-RAID<->GigEthernet with encryption, 90MByte/s without, on modern
> 2 - 3GHz PeeCees. In that case it's just decryption rather than
> ssh+sshd running on the same CPU, so divide that throughput in half,
> and you are in the same MB per MHz ballpark as the other results. I
> think it is probably working properly.
I think you are right. Getting about 1MB/s on a 300MHz UltraSPARC II with
gcc seems pretty good, actually. I figured I would get about half the disk
rate, so about 7MB/s, give or take. Clearly that will not happen.
Sorry that I can't keep the thread going, but I don't subscribe to the list,
and I've not been copied on the messages.
Frank
More information about the freebsd-sparc64
mailing list