[squid-users] Re: What are recommended settings for optimal sharing of cache between SMP workers?

From: Rajiv Desai <rajiv_at_maginatics.com>
Date: Tue, 18 Feb 2014 02:42:47 -0800

Some more info:

Following are mgr:storedir stats after back to back downloads for 4 GB
data (ie same 2 GB twice).
Perhaps the 477 StoreEntries with MemObjects AND 468 Hot Object
Cache Items are not shared?

<stats>
Connection information for squid:

Number of clients accessing cache: 10

Number of HTTP requests received: 78410

Number of ICP messages received: 0

Number of ICP messages sent: 0

Number of queued ICP replies: 0

Number of HTCP messages received: 0

Number of HTCP messages sent: 0

Request failure ratio: 0.00

Average HTTP requests per minute since start: 8167.3

Average ICP messages per minute since start: 0.0

Select loop called: 1228150 times, 4.226 ms avg

Cache information for squid:

Hits as % of all requests: 5min: 44.7%, 60min: 44.7%

Hits as % of bytes sent: 5min: 44.9%, 60min: 44.9%

Memory hits as % of hit requests: 5min: 0.0%, 60min: 0.0%

Disk hits as % of hit requests: 5min: 88.9%, 60min: 88.9%

Storage Swap size: 2454128 KB

Storage Swap capacity: 1.2% used, 98.8% free

Storage Mem size: 1980 KB

Storage Mem capacity: 0.0% used, 0.0% free

Mean Object Size: 62.81 KB

Requests given to unlinkd: 0

Median Service Times (seconds) 5 min 60 min:

HTTP Requests (All): 0.03394 0.03394

Cache Misses: 0.04430 0.04430

Cache Hits: 0.02041 0.02041

Near Hits: 0.00000 0.00000

Not-Modified Replies: 0.00000 0.00000

DNS Lookups: 0.02896 0.02896

ICP Queries: 0.00000 0.00000

Resource usage for squid:

UP Time: 576.032 seconds

CPU Time: 476.166 seconds

CPU Usage: 82.66%

CPU Usage, 5 minute avg: 118.39%

CPU Usage, 60 minute avg: 88.17%

Process Data Segment Size via sbrk(): 81444 KB

Maximum Resident Size: 4967328 KB

Page faults with physical i/o: 2

Memory usage for squid via mallinfo():

Total space in arena: 82632 KB

Ordinary blocks: 75092 KB 4345 blks

Small blocks: 0 KB 0 blks

Holding blocks: 350592 KB 73 blks

Free Small blocks: 0 KB

Free Ordinary blocks: 7540 KB

Total in use: 7540 KB 2%

Total free: 7540 KB 2%

Total size: 433224 KB

Memory accounted for:

Total accounted: 27358 KB 6%

memPool accounted: 27358 KB 6%

memPool unaccounted: 405866 KB 94%

memPoolAlloc calls: 17695559

memPoolFree calls: 17764839

File descriptor usage for squid:

Maximum number of file descriptors: 589824

Largest file desc currently in use: 35

Number of file desc currently in use: 135

Files queued for open: 0

Available number of file descriptors: 589689

Reserved number of file descriptors: 900

Store Disk files open: 1

Internal Data Structures:

  477 StoreEntries

  477 StoreEntries with MemObjects

  468 Hot Object Cache Items

39070 on-disk objects
</stats>

On Tue, Feb 18, 2014 at 1:52 AM, Rajiv Desai <rajiv_at_maginatics.com> wrote:
> Hello,
>
> Need some guidance for optimal sharing of cache amongst SMP workers
> using Large rock.
>
> Context:
>
> I am using squid cache with 8 SMP workers and a 200 GB rock cache stored on SMP.
> (Using squid-3.HEAD-20140127-r13248 which has LargeRock support to
> cache objects > 32 KB).
>
> I have set :
> maximum_object_size 4 MB
> cache_dir rock /mnt/squid-cache 204800 max-size=4194304
> cache_mem 0 MB
>
> My refresh pattern is very permissive which basically allows caching everything:
> refresh_pattern . 129600 100% 129600 ignore-auth
>
>
> Questions:
>
> I am trying to test optimal caching behavior where after I have
> downloaded 1 GB data (~16000 objects ~64KB each) the subsequent
> download should all be cache hits.
> However, I see several TCP misses in access.log. The hit ratio on
> subsequent download of the same objects is ~85%.
>
> 1. Is this expected in a multiple worker setup? If yes, can you please
> briefly explain what contributes towards these misses and how can this
> be minimized?
>
> 2. I set cache_mem to 0 coz I believe cache_mem sharing is
> opportunistic and hence all workers might not always share all objects
> in cache_mem. Is that correct? What impact does cache_mem size have on
> deterministic sharing of cache between workers.
>
> 3. Moreover are there any tips/guidelines to optimize % hit rate for
> previously downloaded objects.
>
> Thanks,
> Rajiv
Received on Tue Feb 18 2014 - 10:42:56 MST

This archive was generated by hypermail 2.2.0 : Tue Feb 18 2014 - 12:00:06 MST