[geeks] Flash drive questions

[Mike Meredith]

On Mon, 7 Aug 2006 19:15:28 +0100, Mike Meredith wrote:
> On Sun, 6 Aug 2006 23:36:10 -0400, Charles Shannon Hendrix wrote:
> > Seek time is not all of paging.  Lot's of paging is big chunks.
> 
> Well sure.

I got curious, and did a little measurement. I've used a dtrace script
on a moderately busy 6900 with 8 cores running a variety of stuff
including Oracle and ran the output through a simple awk script to
pick up pagein activity :-

#    Size
----------
1049 8192
 108 16384
  33 24576
  17 2048
  14 40960
  14 1048576
  12 32768
   7 49152
   6 90112
   5 81920
   3 57344
   3 1024

(I excluded from the list those page sizes below a count of 3)

Whilst (with Solaris) you can get large pageins, the activity is
dominated in number of pages by 8192 byte pages.

Now with an mSystems mSSD Ultra320 you have an 0.02ms seek time, and
with the Seagate ST373455LW you have an 'average' seek time of 3.5ms
(I'm not overly impressed with just an 'average' but that's all they
give you). That lets the mSSD device read for 3.3ms before the Seagate
gets to the data it wants (in the average case). 

With a read speed of 40MBps, that's 40 x 0.0033 Mbytes or 40 x 0.0033 x
1024 Kbytes. Or 135Kbytes. Which means (unless I've made a stupid
mistake) that the slower mSSD device can service the majority of pagein
requests quicker than the Seagate.

Now with large pagein requests the Seagate will win, so you don't want
those files on the mSSD but the dtrace script will tell me the filenames
that would benefit.

Of course I doubt this would make a huge difference in overall
performance for most machines. Although I wouldn't mind trying it on a
V1280 of mine that has *lots* of filesystem pageins.