[squid-users] Squid 3.5.21 - High CPU (100%)

Jasper Van Der Westhuizen jvdwesthuiz at shoprite.co.za
Tue Sep 27 04:06:23 UTC 2016




> Is there anything that you guys can suggest I do around the cache?
> Should I try a different store type? A different filesystem type
> perhaps?
>
>
>
> If your store has a configuration knob that effectively limits disk
> writing rate, then use it to limit that rate to avoid overflowing the
> queue.
>
> You can also consider using rock store that has an explicit write
> rate limiting option (but comes with other problems that may or may
> not affect your setup).
>
> Adding more physical disk spindles helps, of course.
>

Just to be clear since you mentioned working with virtual services. By
"spindle" Alex refers to individual physical HDD devices underneath the
whole VM storage setup. 'Disk' a the VM level, even at the RAID level
might refer to multiple or overlapping (shared) 'spindles'.

Unless your AUFS/UFS/diskd cache_dir are limited to being on one
"spindle" HDD, they can cause a shared HDD controller to fill up /
overload and adding more cache_dir just makes that particular problem worse.

Amos


Hi Amos

At the time of my email my cache_dir was a 100GB disk assigned from the same datastore as where the VM is. I have to say that I never had problems with this setup. The storage is provided from a IBM XIV storage unit so it's not the slowest of storage devices. After some discussions with a couple of people we thought that perhaps the cache is too big, so I now made it 4 x 15GB aufs cach_dirs. 2 Disks are from the same datastore as the VM, and 2 are from a datastore assigned to the cluster from a SVC (XIV behind it) with the following in my squid.conf.

maximum_object_size_in_memory 256 KB
maximum_object_size 4096 MB
store_dir_select_algorithm least-load

cache_dir aufs /var/cache/squid-c1/ 10240 32 256
cache_dir aufs /var/cache/squid-c2/ 10240 32 256
cache_dir aufs /var/cache/squid-c3/ 10240 32 256
cache_dir aufs /var/cache/squid-c4/ 10240 32 256
cache_replacement_policy heap lfuda
memory_replacement_policy lru


Do you think this is optimal?














Disclaimer:
http://www.shopriteholdings.co.za/Pages/ShopriteE-mailDisclaimer.aspx
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.squid-cache.org/pipermail/squid-users/attachments/20160927/17087f22/attachment.html>


More information about the squid-users mailing list