Re: So I tried POV myself (oops)

[Jeff Licquia <jeff@web.lanscape.net>]

On Tue, 17 Mar 1998, Erich Schroeder wrote:

> 
> Oops, I hadn't done the command correctly. Running the correct benchmark 
> command I get:
> 
> Time For Trace:    0 hours  8 minutes   6.0 seconds (486 seconds)
>     Total Time:    0 hours  8 minutes   6.0 seconds (486 seconds)
> 472.34user 0.83system 8:07.76elapsed 97%CPU (0avgtext+0avgdata 0maxresident)k
> 0inputs+0outputs (0major+0minor)pagefaults 0swaps
> 
> the first time and:
> 
> Time For Trace:    0 hours  7 minutes  39.0 seconds (459 seconds)
>     Total Time:    0 hours  7 minutes  40.0 seconds (460 seconds)
> 456.68user 0.46system 7:40.72elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
> 0inputs+0outputs (0major+0minor)pagefaults 0swaps

Yes, those numbers make more sense.

I had the same problem when I ran the benchmarks for the first time.  It 
ran in about 30 seconds, and I thought "whoa!"

> For the second time. Again, what's the deal with the differences? Still, 
> these numbers make much more sense than my first ones.

Part of the difference is disk cache.  The first time, the system has to 
load the file itself, as well as several included files, from disk, 
wasting precious seconds; the second and following times, those files 
will be read from cache.

I was willing to explain the whole 15 second difference in the "Lucy" run 
with that, since the caching problem would affect all nodes in the 
cluster and network latency would also be a factor, but I'm not sure I 
would count that as the factor for all 25 seconds of your differences.  
Maybe processor cache?  (I'm just guessing now.)

What would be interesting to know is whether the speedup would be 
attained if you interrupted the task after, say, 10 seconds and re-ran 
it.


--
To unsubscribe, send email to majordomo@luci.org with
"unsubscribe luci-discuss" in the body.