Alex,
Thanks for the clarification, although I'm now going to ask for more :)
I'm running a slightly modified polymix-3 test (I took out the framp and
fexit phases entirely) to see how long it takes to actually fill my 10 GB
cache. I specified a fill rate of 200 req/sec, but the load that is being
generated hangs around 15 to 25 req/sec. I understand what you've said
below about only the "fill" responses, but I'm not even getting the request
rate that I requested. I expected to see my req/sec jump right up near 200
req/sec and continue at that rate until the "fill" responses hit the
2*cachesize mark. I'm using polygraph 2.5.4.
Here are the appropriate lines from my modified polymix-3.pg file.
TheBench.peak_req_rate = 200/sec;
size ProxyCacheSize = 10GB;
rate FillRate = 200/sec;
And here is some sample console output:
247.69| i-fill 298462 21.60 2978 0.00 0 1666
247.77| i-fill 298576 22.80 2642 6.48 0 1656
247.86| i-fill 298700 24.80 2785 0.00 0 1653
247.94| i-fill 298815 23.00 2462 14.41 0 1654
248.02| i-fill 298897 16.40 2874 2.47 0 1656
Did taking the "framp" stage out of polymix-3-guts.pg file affect this?
Brian O'Keefe (brianok@us.ibm.com)
IBM Transarc Lab, Pittsburgh, PA
Alex Rousskov <rousskov@ircache.net> on 08/03/2000 09:58:38 AM
Brian,
PolyMix rules require to "fill" your cache twice. That is, the
fill goal is twice the configured cache size. Moreover, not all
responses are "fill" responses. Only basic cachable misses are counted
as "fill" responses. Thus, the actual fill rate is smaller than
configured request rate.
IIRC, average object size is about 11KB. For the actual contribution of
fill transactions and average fill object size, use Report Generator.
Hint: Reduce the cache size in polymix-3.pg to, say, 100MB and reduce
plateau duration to, say, 5 minutes. Run that short test and generate a
report _for the fill phase_. The statistics will tell you fill
contribution and object sizes without waiting for hours...
Alex.
P.S. Polygraph 2.5.4 has some tweaks to speedup the fill (it does only
basic requests during the fill phase).
On Thu, 3 Aug 2000 brianok@us.ibm.com wrote:
>
>
> I'm trying to understand how the cache fill takes place for PolyMix-3.
Does
> it just assume that the cache gets filled based on the amount of data
that
> is sent during the i-fill phase? For example, if I specify a 10 GB cache,
> and a 200 req/sec fill rate, and assuming an average request size to be
10
> KB (empirically, it seems to be more like in the teens: 13KB - 18KB), it
> should take about 83 minutes to fill my cache. Is this correct? Is that
was
> PolyMix-3 uses roughly to determine when to move on to i-fexit and
i-inc1?
>
> I'm asking because my filling is taking much longer than that, and in
fact
> the polyclt doesn't get out of i-fill before it fails due to the "Virtual
> memory exceeded in 'new'" message (see my previous post). Maybe these two
> problems are related? The time it is taking to fill my cache leads me to
> believe that there is something wrong here. Thanks for any help.
>
> Brian O'Keefe (brianok@us.ibm.com)
> IBM Transarc Lab, Pittsburgh, PA
This archive was generated by hypermail 2b29 : Tue Jul 10 2001 - 12:00:14 MDT