sön 2006-03-12 klockan 14:59 +0000 skrev Michael Pye: > Some analysis of squid (2.5stable12, compiled using sun studio cc) using > dtrace on solaris10 shows in my acclerator setup that there are large > numbers of READ system calls apparently to small files in the cache > directory. These are hits to images that clients have and are sending > If-Modified-Since requests, so the log shows lots of TCP_IMS_HIT accesses. Ok. > The amount of reads to these small files though indicates that these > files are read everytime. Correct, and is one reason why noatime is a good mount option for Squid caches.. > Is there not a way to get these into memory so > the TCP_IMS_HITs can be answered from memory rather than from disk? By forcing a reload of the objects. Only data fetched from the network ends up in the memory cache (the memory cache is simply speaking that data received from network is kept in an internal receive buffer, same buffer space as used for in-transit objects...). > I still have minimum_object_size to 0 and a large enough > maximum_object_size for this to be in memory. These two are not memory related. Perhaps you were thinking of maximum_object_size_in_memory? > In fact the object is in > memory as clients that don't send IMS requests do get the object via a > TCP_MEM_HIT. Is there some reason why IMS requests are not answered from > memory ? Hmm.. are you absolutely sure about this? My Squid does not show such symptoms.. Note: The logs does not make any difference between IMS handled from memory or from disk. Both are logged as TCP_IMS_HIT. Regards Henrik
Attachment:
signature.asc
Description: Detta =?ISO-8859-1?Q?=E4r?= en digitalt signerad meddelandedel
