Christian Robottom Reis wrote:
>
> James, I've run a whole truckload of benchmarks on raid1 with varying
> chunksizes on three different kernels, and on a plain disk. I'm about to
> publish some of the stuff, but I'm wondering very hard why is it that the
> readbalancing test showed _awful_ numbers on tiotest 0.21 and great
> numbers on 0.24 - any idea? Just have a look:
>
> [this is big - my apologies]
>
> Chunk is 4k, Stride is 4 and Block is 1024k
> Found memory size of 48.00390625 MB
> Size is MB, BlkSz is Bytes, Read and Write are MB/sec, Seeks are Seeks/sec
>
> Machine Directory Size(MB) BlkSz Threads Read Write Seeks
> ----------- ----------- --------- ------- --------- ------ ------- -------
> /bench 512 4096 1 16.258 10.3448 112.341
> /bench 512 4096 2 13.798 10.7370 101.831
> /bench 512 4096 4 4.9674 11.0868 92.2756
> /bench 512 4096 8 4.6620 11.9793 97.2662
> /bench 512 4096 16 4.8134 13.5368 99.7289
> /bench 512 4096 32 5.1743 12.9215 113.547
> /bench 512 4096 64 5.1585 12.3916 117.426
> Chunk is 4k, Stride is 2 and Block is 2048k
> Found memory size of 48.00390625 MB
> Size is MB, BlkSz is Bytes, Read and Write are MB/sec, Seeks are Seeks/sec
>
Great, real numbers are always worth a look. Anyone want to help me
figure out how to interpret these numbers?
1. I assume that the first two, threads 1 and 2, are showing the effect
of DRAM to some extent or the accesses are more sequential. But,
there is still the relatively high seeks/sec, so I go with DRAM.
2. I guess that the seeks/sec are per drive, but that just scales the
numbers a little differently. Looks like transfers are in the
range of 64KB.
3. A ratio of ~1:2 for Reads to Writes seems ok if the driver is not
optimizing read accesses.
4. So, as the number of threads increases, caching effects decrease,
queuing improves and then system load slows throughput?
Example:
> Threads Read Write Seeks Rd+Wr (Rd+Wr)/Sk
> --------- ------ ------- ------- ------- ----------
> 1 16.258 10.3448 112.341 25.6 (230/2)KB Memory?
> 2 13.798 10.7370 101.831 24.5 (240/2)KB Memory?
> 4 4.9674 11.0868 92.2756 16.0 (170/2)KB
> 8 4.6620 11.9793 97.2662 16.6 (170/2)KB Queuing?
> 16 4.8134 13.5368 99.7289 18.3 (180/2)KB Queuing?
> 32 5.1743 12.9215 113.547 18.1 (160/2)KB
> 64 5.1585 12.3916 117.426 17.5 (150/2)KB uP loading?
--
Dan Jones, Storage Engineer VA Linux Systems
V:(408)542-5737 F:(408)745-9130 1382 Bordeaux Drive
[EMAIL PROTECTED] Sunnyvale, CA 94089
