System Software Telecon Minutes

[Jim Gettys]

On Tue, 2006-11-07 at 17:12 -1000, Mitch Bradley wrote:

> Chris tested NAND, it worked well.  Chris thinks NAND is 3x the speed
> of SD under load.  3 MiB/sec vs 1 MiB.  We think we have used the
> SD at high speed mode - 48 MHz seen on scope.  Patches for HS mode not
> yet in kernel.  That needs to happen soon.

Just so people understand: Chris Ball was running dbench, a multiprocess
file system exerciser (the intent is to stress the device driver); so
the performance numbers for SD are when there are many readers/writers,
which would be very unusual in our use patterns, and the performance is
for just one of the N processes running at once (the aggregate
performance of both NAND and SD is much higher).

With the much higher latency of SD relative to bare NAND, it having
lower single stream performance (this number is from one of a set of
processes) is hardly surprising.

We need to take some data on single stream throughput on SD: there the
numbers were much more comparable between SD and NAND, as you would
expect, but we didn't do the experiments carefully and should retake
them.

Neither SD or NAND will be able to run at full speed until we have CaFE
back as an asic: the FPGA cannot run at anything like the speed the ASIC
will run once we have it.
                           - Jim


-- 
Jim Gettys
One Laptop Per Child