Search squid archive

Re: Tune Squid proxy to handle 90k connection

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



I would recommend you to start with 0 caching.
However, for choosing the right solution you must give more details.
For example there is an IBM reasearch that prooved that for about 90k connections you can use vm's ontop of such hardware with apache web server.
If you do have the set of the other requirements from the proxy else then the 90k requests it would be wise to mention them.

Do you need any specific acls?
Do you need authentication?
etc..

For a simple forward proxy I would suggest to use a simpler solution and if possible to not log anything as a starter point.
Any local disk i/o will slow down the machine.

About the url categorization, I do not have experience with ufdbguard on such scale but it would be pretty heavy for any software to handle 90k rps...
 It's doable to implement such setup but will require testing.
Will you use ssl bump in this setup?

If I will have all the technical and specs/requirements details I might be able to suggest better then now.
Take into account that each squid worker can handle about 3k rps tops(with my experience) and it's a juggling between two sides so... 3k is really 3k+3k+external_acls+dns...

I believe that in this case an example of configuration from the squid developers might be usefull.

Eliezer


בתאריך יום ג׳, 25 בינו׳ 2022, 18:42, מאת André Bolinhas ‏<andre.bolinhas@xxxxxxxxxxxxxx>:
Any tip about my last comment?

-----Mensagem original-----
De: André Bolinhas <andre.bolinhas@xxxxxxxxxxxxxx>
Enviada: 21 de janeiro de 2022 16:36
Para: 'Amos Jeffries' <squid3@xxxxxxxxxxxxx>; squid-users@xxxxxxxxxxxxxxxxxxxxx
Assunto: RE: Tune Squid proxy to handle 90k connection

Thanks Amos
Yes, you are right, I will put a second box with HaProxy in front to balance the traffic.
About the sockets I can't double it because is a physical machine, do you think disable hyperthreading from bios will help, because we have other services inside the box that works in multi-threading, like unbound DNS?

Just more a few questions:
1º The server have 92Gb of Ram, do you think that is needed that adding swap will help squid performance?
2º Right now we are using squid 4.17 did you recommend upgrade or downgrade to any specific version?
3º We need categorization, for this we are using an external helper to achieve it, do you recommend use this approach with ACL or move to some kind of ufdbguard service?

Best regards
-----Mensagem original-----
De: squid-users <squid-users-bounces@xxxxxxxxxxxxxxxxxxxxx> Em Nome De Amos Jeffries
Enviada: 21 de janeiro de 2022 16:05
Para: squid-users@xxxxxxxxxxxxxxxxxxxxx
Assunto: Re: Tune Squid proxy to handle 90k connection

Sorry for the slow reply. Responses inline.


On 14/01/22 05:44, André Bolinhas wrote:
> Hi
> ~80k request per second  10k users


Test this, but you may need a second machine to achieve the full 80k RPS.

Latest Squid do not have any details analysis, but older Squid-3.5 were only achieving >15k RPS under lab conditions, more likely expect under 10k RPS/worker on real traffic.
  That means (IME) this machine is quite likely to hit its capacity somewhere under 70k RPS.


> CPU info:
> CPU(s) 16
> Threads per code 2
> Cores per socket 8

With this CPU you will be able to run 7 workers. Setup affinity of one core per worker (the "kidN" processes of Squid). Leaving one core to the OS and additional processing needs - this matters at peak loading.

CPU "threads" tend not to be useful for Squid. Under high loads Squid workers will consume all available cycles on their core, not leaving any for the fancy "thread" core sharing features to pretend there is another core available. YMMV. One of the tests to try when tuning is to turn off the CPU hyperthreading and see what effect it has (if any).


> Sockets 1
> Inter Xeron Silver 4208  @ 2.10GHz
>

Okay. Doable, but for best performance you want as high GHz rating on the cores as your budget can afford. The amount of "lag" Squid adds to traffic and RPS performance/parallelism directly correlates with how fast the CPU core can run cycles.



HTH
Amos
_______________________________________________
squid-users mailing list
squid-users@xxxxxxxxxxxxxxxxxxxxx
http://lists.squid-cache.org/listinfo/squid-users

_______________________________________________
squid-users mailing list
squid-users@xxxxxxxxxxxxxxxxxxxxx
http://lists.squid-cache.org/listinfo/squid-users
_______________________________________________
squid-users mailing list
squid-users@xxxxxxxxxxxxxxxxxxxxx
http://lists.squid-cache.org/listinfo/squid-users

[Index of Archives]     [Linux Audio Users]     [Samba]     [Big List of Linux Books]     [Linux USB]     [Yosemite News]

  Powered by Linux