On 7/3/20 4:50 AM, patrick mkhael wrote:
> workers 3
> cpu_affinity_map process_numbers=1,2,3,4,5,6 cores=1,2,3,4,5,6
> cache_dir rock /rock1 200000 max-size=32000 swap-timeout=300 max-swap-rate=100
> cache_dir rock /rock2 200000 max-size=32000 max-swap-rate=100 swap-timeout=300
> cache_dir rock /rock3 200000 max-size=32000 max-swap-rate=100 swap-timeout=300
> cache_mem 17 GB
> maximum_object_size_in_memory 25 MB
> maximum_object_size 1 GB
> cache_miss_revalidate off
> quick_abort_pct 95
FYI: The combination of 1GB maximum_object_size and much smaller size
limits for objects in memory and disk caches does not make sense: There
is no cache to store a, say, 26MB object. If Squid lacks the
corresponding configuration "lint" check, somebody should add it.
> This config is giving 4% of cache gain ratio,
> in addition as i already mentionned before if i take the same above
> config without worker and cach_dir with the same traffiic using aufs on
> one of the disks , i have automatically i har 60 % cache ratio.
When using AUFS, do you limit disk-cached object sizes to 32KB like you
do with rock? If not, then you should remove the max-size limit from
rock cache_dirs. Modern rock cache_dirs are capable of storing large
objects.
What kind of hit ratio do you get with rock if you do not limit
swap-rate and do not specify swap-timeout?
What kind of hit ratio do you get with rock if you use one worker, one
rock cache_dir, do not limit swap-rate, do not specify swap-timeout, and
start Squid with -N to disable SMP?
As you can see, I am trying to understand whether the size limitation,
the rate limiting, or SMP problems explain the drop in hit ratio.
> Shoud rock give me the same performance as aufs ?
It is a difficult question to answer correctly (for me). The goal is for
rock performance to exceed that of (a)ufs, but I doubt we have reached
that goal in every environment that matters (including yours).
* In a non-SMP environment, I would expect similar hit ratios in most
cases, but I would not be surprised if there are significant exceptions.
Rock is focused on SMP support, and there are complexities/costs
associated with SMP. Rock is getting better, but there are some known
areas where rock cannot yet do what ufs (including aufs) can.
* In a SMP environment, the question is mostly meaningless because there
is no SMP support for ufs-based caches. Folks use squid.conf
preprocessor hacks to configure ufs-based caches in SMP mode, but those
setups usually violate HTTP and may cause serious problems. YMMV.
> for a traffic of 1 Gb/s , is there a way to use aufs ?
Before trying unsupported combinations of aufs and SMP, I would try to
understand why your hit ratio is so low with rock. The questions above
may be a good start in that investigation.
Cheers,
Alex.
> ------------------------------------------------------------------------
> *From:* Alex Rousskov <rousskov@xxxxxxxxxxxxxxxxxxxxxxx>
> *Sent:* Thursday, July 2, 2020 4:24 PM
> *To:* patrick mkhael <patrick.mkhael@xxxxxxxxxxx>;
> squid-users@xxxxxxxxxxxxxxxxxxxxx <squid-users@xxxxxxxxxxxxxxxxxxxxx>
> *Subject:* Re: rock issue
>
> On 7/1/20 4:45 PM, patrick mkhael wrote:
>
>> ***Please note that you have 20 kids worth mapping (10 workers and 10
>> diskers), but you map only the first 10.{since i did not get the point
>> of the diskers ,as far as i understood , it should be like (simple
>> example)
>
>>> workers 2
>>> cpu_affinity_map process_numbers=1,2,3,4 cores=1,2,3,4
>>> cache_dir rock ...
>>> cache_dir rock ...
>
> The above looks OK. Each worker is a kid process. Each rock cache_dir is
> a kid process (we call them diskers). If you have physical CPU cores to
> spare, give each kid process its own physical core. Otherwise, give each
> worker process its own physical core (if you can). Diskers can share
> physical cores with less harm because they usually do not consume much
> CPU cycles. Squid wiki has more detailed information about that:
> https://wiki.squid-cache.org/Features/SmpScale#How_to_configure_SMP_Squid_for_top_performance.3F
>
>
>> ***Why do you have 10 rock caches of various sizes? [ to be honest , i
>> saw in many websites that it should be like this from the smallest to
>> the bigest with diff size, i tought it should serve from small size pool
>> to high ]
>
> IMHO, you should stop reading those web pages :-). There is no general
> need to segregate objects by sizes, especially when you are using the
> same slot size for all cache_dirs. Such segregation may be necessary in
> some special cases, but we have not yet established that your case is
> special.
>
>
>> *****How many independent disk spindles (or equivalent) do you have? [ i
>> have one raid 5 ssd disks , used by the 10 rock cache dir]
>
> Do not use RAID. If possible, use one rock cache_dir per SSD disk. The
> only reason this may not be possible, AFAICT, is if you want to cache
> more (per SSD disk) than a single Squid cache_dir can hold, but I would
> not worry about overcoming that limit at the beginning. If you want to
> know more about the limit, look for "33554431" in
> http://www.squid-cache.org/mail-archive/squid-users/201312/0034.html
>
>
>> ***How did you select the swap rate limits and timeouts for
>> cache_dirs?[I took it also from online forum , can i leave it empty for
>> both]
>
> If you want a simple answer, then it is "yes". Unfortunately, there is
> no simple correct answer to that question. To understand what is going
> on and how to tune things, I recommend studying the Performance Tuning
> section of https://wiki.squid-cache.org/Features/RockStore
>
>
>> ****Do you see any ERRORs or WARNINGs in cache log?[NO error or warning
>> found in cache]
>
> Good. I assume you do see some regular messages in cache.log. Keep an
> eye for ERRORs and WARNINGs as you change settings.
>
>
> HTH,
>
> Alex.
_______________________________________________
squid-users mailing list
squid-users@xxxxxxxxxxxxxxxxxxxxx
http://lists.squid-cache.org/listinfo/squid-users
