On 10/25/2013 06:37 AM, Ahmad wrote: > im trying to use smp feature on squid last version 3.3.9 compiled , > i have server of Delr210 with 8 G rams and i have cpu quad core 3450 intel > xeon , > so , > my loads insquid is as below : > 1-disk cahing "i have 4 hardsisks each with 90 G cache dir"" > 2-memory caching " i have 8 giga memory installed " > 3-web filtering "using ACLS" > 4-custom ACLS based on srx ips > > > i read http://wiki.squid-cache.org/ConfigExamples/SmpCarpCluster Do you need CARP? If you do not need it or are not sure, then I recommend starting with a single SMP Squid instance (multiple workers, one squid.conf) as discussed below. > Q1-how many workers do i need ? " from /proc/meminfo i have 8 cores " Please note that you do not have 8 physical cores. You have 4. You have 8 virtual (a.k.a., hyperthreaded) cores that are not very useful for busy Squid workers (they only waste cycles on resource contention if two busy Squid workers share the same physical core). > Q2- what process is better to go frontend and which loads better to be in > backend ?? Assuming you want to optimize performance, you can try the following mapping as a starting point (using hyperthreaded cores, and assuming that your core #1 and #2 share the same physical core -- something you need to verify because some CPUs have different virtual:real core mapping!): Core #1 - Left for OS, NICs, etc. Bind NIC to this core. Core #2 - Rock disker for cache_dir #1. Core #3 - Worker #1 Core #4 - Rock disker for cache_dir #2. Core #5 - Worker #2 Core #6 - Rock disker for cache_dir #3. Core #7 - Worker #3 Core #8 - Rock disker for cache_dir #4. If your tests (or live deployment) show that workers are overloaded, then you can try adding Worker #4 on Core #2. If that is not enough, you would need a beefier server. If your tests (or live deployment) show that workers and diskers compete for CPU cycles too much, then you would need to reduce disk caching or get a beefier server. Other adjustments may be necessary, of course -- the above is just a starting point. HTH, Alex.
