Re: Interesting speed benchmarks

In message: <45BA51EA.3070901_at_freebsd.org>
            Eric Anderson <anderson_at_freebsd.org> writes:
: On 01/26/07 12:49, Warner Losh wrote:
: > From: Stefan Ehmann <shoesoft_at_gmx.net>
: > Subject: Re: Interesting speed benchmarks
: > Date: Fri, 26 Jan 2007 10:52:11 +0100
: > 
: >> On Friday 26 January 2007 03:24, M. Warner Losh wrote:
: >>> On a lark, I just got a combo USB/Firewire external disk drive.  I ran
: >>> some crude benchmarks, and I was surprised by what I found.  This is
: >>> on a fairly stock -current kernel.
: >>>
: >>> Firewire does around 40MB/s, while USB 2.0 maxes out at about 12MB/s.
: >>> This is with a simple dd command:
: >> On my i386 notebook with USB 2.0 enclosure.
: >> Linux: 31.5MB/s
: >> FreeBSD: 27.5MB/s
: >>
: >> There's still room for improvement but numbers don't seem that bad.
: >>
: >> Maybe you should try knoppix or so to verify it's not the drive's fault. Other 
: >> than that I'd also guess it's an amd64 problem.
: > 
: > It is not an AMD64 problem.  I get the same numbers on my i386 latpop
: > as I get on my amd64 laptop.  Actually, I get WORSE numbers on the
: > i386 laptop by about 20%.
: > 
: > It isn't the drive's fault.  Otherwise, firewire wouldn't get 40MB/s.
: > The same drive, the same enclusure are used for both the USB and
: > firewire tests.  It is about as apples to apples as you can get.
: > 
: > There's some serious performance issues in the usb stack.
: > 
: > Warner
: 
: 
: A few tidbits of information (may be useful, maybe not):
: 
: - I've seen the firewire part of the enclosure be faster than the USB 
: part.  The chips that run it are possibly different, so that shouldn't 
: be forgotten.  I've had a few USB->flash adapters that got lousy 
: performance, but when I switched to a USB->SATA flash card reader, the 
: performance doubled.

I'm now seeing on my FreeBSD desktop at work numbers that are in the
28MB/s range.  I'll have to investigate things more closely...  I
can't imagine why FreeBSD/amd64 would be so much slower.

: - For those testing using a file system - STOP! It's not a good test of 
: the throughput of the device, and depends on a lot of variables.  dd or 
: diskinfo are decent generic tools, but in Windows you just can't use a 
: file system benchmark to compare.

yes.

: - If you read/write less than the drive cache, it should remove the 
: latency of the drive from the equation, right?

True, but not relevant, I don't think.  the speed of the disk is in
excess of what either firewire or usb can do.

Warner