On 07/18/2013 06:10 AM, Travel Factory S.r.l. wrote: > > Yesteray I moved for several hours all my users to the 2 new servers. > > Since I want to test SMP / Rock and eventually SMPCarp I went to have a > look at my logs. > > My first goal is understand which max-size to set to rock cache_dir. > So I did this on one server: > grep SWAPOUT store.log.0 | sed -e 's/ */|/g' | cut -d"|" -f11| cut > -d"/" -f2 > sizes > wc -l sizes > 491458 > followed by: > > cat sizes | sort -g | uniq -c > result > wc -l result > 68900 > > you can download the file, if you want, from www.bruxx.it/frank/result. > > These are SWAPOUT entries and as far as I know they are stored on > disk... are they ? > > You will notice that there are 18902 requests for 43 bytes SWAPOUT. > 15414 are from http://p.twitter.com/t.gif? > > Is it normal that these files are cached ? > > 1374048878.108 SWAPOUT 00 00008061 FE66CED6D9B9E3E31D39654ED9FE19FA 200 > 1374048878 1328738114 -1 image/gif 43/43 GET > http://p.twitter.com/t.gif? > > I can't find a single HIT in access log, well.... ok, I have 1265 > TCP_REFRESH_UNMODIFIED/304 > On my prodution server (squid 2.7) I only have TCP_MISS in the logs ! > > > So I arrive at the questions: > > Is it normal that these queries, with the ?, are cached ? > > Is there a list of domains/pages that it is better not to cache since > they are changing anyway ? > > After removing the 3213 entries with 0 bytes, I have 302541 entries with > less than 9000 bytes... they cover 75% of the cached requests... is 9000 > a good tradeoff ? IMHO, you may be overthinking this. Current rock storage cannot store objects larger than 32KB, the maximum slot size. With modern disk drives, you can use that as the slot size for your rock store and still cache lots of objects. 32KB covers the vast majority of responses so you have a good chance of arriving at a well-working configuration. If everything works OK, great. If not, if you want to experiment, or if you really want (and can) cache more objects, then you can try reducing the rock slot size by half to 16KB and compare the results. HTH, Alex.
