In article <guycoleCDBJ5w.HGB@netcom.com> guycole@netcom.com (Guy Cole) writes:
>
>3. There is a competing product from Auspex, which is supposed to be
>quite good.
We bought Auspex' own write accelerator board, which has a 1MB NVRAM cache,
for our NS/5000, and below is the performance improvement we saw in NFS
response times (in ms). Response times were measured with NetMetrix running
on a Sun, and in both cases there were well over two million requests during 
the measurement period. I would like to stress that this was observed on a 
normal work day; there were no special tests or benchmarks used. Operations 
are listed in order of frequency; a typical mix would be getattr 30%, lookup 
21%, readlink 18%, read 12%, write 11%, readdir 3%, all others < 1% each. 
[Under some circumstances (i.e. people deciding to store large quantities of
data on the server and process it, which we normally discourage), writes can
be as much as 20% of the mix over a period of several days. So having the
accelerator is definitely a good thing for us.]
      NFS Operation         Before            After
        getattr              18.7              2.6  (retransmits way down too)
         lookup               7.4              6.6
       readlink               4.6              3.5
           read              17.3             19.8
          write             120.9             12.4
        readdir              18.6             16.6
        setattr              24.1              8.0
         create              42.8             20.2
         remove              45.6             15.7
         rename              99.4             25.5
           link              37.8              9.2
          mkdir             117.0             49.5
          rmdir              90.8             90.3
         statfs               4.0              1.5
        symlink              70.0             22.0
"write" benefitted the most, but "getattr" was close behind. We figure that
since getattr is the most common operation, and one of the simplest, it
benefited because the FP didn't have to spend so much time handling all the
other operations that the accelerator helped with, and was able to respond
faster to pretty much everything else as well.
Obviously variations in things like which filesystems were being accessed
the most in the two different runs can account for some of the changes, but 
in most cases the reductions are really significant.
Disclaimer: I have no connection with either Auspex or NetMetrix other than 
as a satisfied customer, I just thought everyone might find this interesting.
-- Ruth Milner NRAO/VLA Socorro NM Computing Division Head rmilner@zia.aoc.nrao.edu
This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:08:11 CDT