A bit of both is probably the most suitable. But it depends on the size distribution of your objects. If mainly small objects then cache_mem is sufficient.
Note: to use more than ca 1-1.5GB of cache_mem you need to build Squid as a 64 bit application. Using tmpfs does not have such limitations for 32 bit applications from what I know. Regards Henrik On Wednesday 19 February 2003 23.52, Omer Shenker wrote: > Hello, > > I'm preparing to running Squid 2.5 as an HTTP accelerator on a > Solaris 8 machine with lots of free memory but a bottleneck in disk > IO. Because I don't want Squid's disk IO needs to preempt those of > PostgreSQL, I'd like to avoid using a disk cache at all. (The > machine is rebooted so infrequently that I don't care about the > lack of permanence of a memory-only cache.) > > My question is this: is it better to just increase the cache_mem > option to the amount of memory I'd like Squid to use, or should I > use TMPFS (SunOS's memory filesystem) as a cache_dir with diskd? > The way I see it, if I went with TMPFS I'd want to turn off > cache_mem completely, otherwise I'd end up with objects cached in > memory twice. Or am I foolish to even consider doing this at all? > > (SunOS admins: I'm aware that I have to make sure /tmp doesn't use > up all of TMPFS and that all TMPFS mounts compete for the same > memory space by default.) > > Unloading services to a different machine is not really an option, > nor is buying another drive. > > Thanks,
