C[omp]ute

Welcome to my blog, which was once a mailing list of the same name and is still generated by mail. Please reply via the "comment" links.

Always interested in offers/projects/new ideas. Eclectic experience in fields like: numerical computing; Python web; Java enterprise; functional languages; GPGPU; SQL databases; etc. Based in Santiago, Chile; telecommute worldwide. CV; email.

Personal Projects

Choochoo Training Diary

Last 100 entries

Raid 5 Speeds

From: "andrew cooke" <andrew@...>

Date: Sat, 7 Apr 2007 09:59:57 -0400 (CLT)

I finally understand (I think) how having raid 5 affects the disk speed.

For large inputs and outputs, and for seeks, it is much faster than single
disks, increasing in speed by a factor of N (or perhaps N-1).  You can see
this by running iostat while the disks are in heavy use.  The IO rates for
the md device are much higher than the disks themselves, and the disks are
limiting the system (lots of IO wait in the CPUs).

This is also clear in bonnie++ output:

Version 1.01d       -Sequential Output- -Seq-Inp- --Random-
                    --Block-- -Rewrite- --Block-- --Seeks--
Machine        Size K/sec %CP K/sec %CP K/sec %CP  /sec %CP
quiet            7G 34986  10 15030   3 43503   4  73.6   0

(reformatted to drop the per character values)

For my cheap disks (320Gb SATA 7200 WD Caviar), those are significant
improvements over a disk in isolation, as you'd expect from striping.

However, "everyone" says how bad raid 5 is for writing, so why are the
write speeds so close to read?  Because, I think, the file being written
is much bigger than the 128K block size, so there's no need to read the
parity information when writing - the system is simply writing a new
parity block along with new data blocks.

This may be reflected in the lower "rewrite" figure, which is dirtying
data within a single block.

Andrew

Rubbish

From: "andrew cooke" <andrew@...>

Date: Sat, 7 Apr 2007 17:35:22 -0400 (CLT)

I think everything I wrote above may be rubbish.  I get the same numbers
on my laptop.  The test used a 4GB file and my current best guess is that
2GB stayed in memory, hence the double speed for read/write over rewrite. 
But really I don't have a clue.  And seeks were actually slightly better
on my laptop.

(The laptop feels slower, but that's more a CPU than a disk thing, I thin
- one reason I have been trying to measure/understand things is because I
sometimes feel there's an odd lag with the server's disks).

Andrew

Iozone Results (Disk Speed Again)

From: "andrew cooke" <andrew@...>

Date: Thu, 12 Apr 2007 21:07:08 -0400 (CLT)

I ran iozone overnight and the results for random reads are shown here -
http://www.acooke.org/cgi/photo.py?start=disk&cols=5&rows=3

I'm not sure how clear that is (I shrank the image a little to fit my
standard "photo" size), but file size is increasing towards the viewer and
throughput drops off a cliff when the file size exceeds memory (the drop
from orange to blue).  The two largest file sizes were 4 and 8 GB.

Along the "foot" of the plot is record size.  There's a slight dip when
record size exceeds about 1MB.  This might be related to the L2 cache (2MB
shared).

iozone - http://www.iozone.org/

Also, gnuplot in X now lets you "drag" 3D plots with the cursor.  Very
cool + useful.

gnuplot - http://www.gnuplot.info/

Andrew

Comment on this post