[squid-users] Scaling concurrent TCP sessions beyond ephemeral port range
Alex Rousskov
rousskov at measurement-factory.com
Sat Sep 10 02:22:02 UTC 2022
On 9/9/22 18:29, Praveen Ponakanti wrote:
>
> On Thu, Sep 8, 2022 at 8:31 PM Alex Rousskov wrote:
>
> On 9/8/22 19:41, Praveen Ponakanti wrote:
> > * We have a large number of workers (30) to help with handling a
> > high RPS. However, TCP session reuse does not seem to be optimal
> > even with server_persistent_connections enabled as a new outbound
> > session would have to be opened up if the request is proxied by a
> > kid worker that doesn’t already have a connection to that
> > destination. Is there something that can be done to improve this
> > with later versions of squid? Would be glad to help out if anyone
> > has some suggestions.
>
> If your only concern is TCP, and the number of servers is large,
> then it
> would be possible to share open Squid-server connections among workers
> by adding code that would exchange open TCP socket descriptors using
> UDS
> messages, but I doubt it is worth doing (a lot of complexity but not
> enough gain). There may also be some advanced/modern kernel tricks that
> we can teach Squid to use for sharing connections, but, again, I doubt
> the complexity would be worth the benefits from such reuse.
> Agree, it might not make sense to increase the complexity with sharing
> socket among the workers. Was thinking more on the lines of a hashmap
> that the coordinator could use to pick workers that already have a TCP
> connection to the destination being requested, instead of having the
> workers themselves share connection details.
Coordinator does not receive/see regular HTTP traffic. If we start
routing HTTP transactions through that process, it may become the
bottleneck itself _and_ will introduce additional overheads for passing
descriptors to workers. From performance point of view, the model with
one "routing" task doling work to workers works best (and is commonly
used) in threaded applications, but Squid is not threaded at that level.
> Most of the TCP connections are for HTTPS reqs, w/o TLS termination at
> the squid. Does squid currently support a TLS session cache ?
Yes, there is some support for worker-specific TLS session caching, with
directives like sslproxy_session_cache_size, tls_outgoing_options
options=NO_TICKET (for outgoing sessions IIRC) and https_port
sslflags=NO_SESSION_REUSE and https_port sslcontext (for incoming sessions).
HTH,
Alex.
> If you are dealing with TLS sessions as well, then we should add a
> shared memory TLS session cache that all workers can tap into.
>
>
> Cheers,
>
> Alex.
>
> > On Tue, Jun 21, 2022 at 2:11 PM Alex Rousskov wrote:
> >
> > On 6/19/22 12:48, Praveen Ponakanti wrote:
> >
> > > What is the process to have this code patch upstreamed for
> future
> > squid
> > > versions?
> >
> > In short, just post a quality pull request on GitHub (or find
> somebody
> > who can guide your code towards official acceptance for you). For
> > details, please see
> https://wiki.squid-cache.org/MergeProcedure
> <https://wiki.squid-cache.org/MergeProcedure>
> > <https://wiki.squid-cache.org/MergeProcedure
> <https://wiki.squid-cache.org/MergeProcedure>>
> >
> >
> > Thank you,
> >
> > Alex.
> >
> >
> > > On Fri, May 20, 2022 at 9:31 PM Amos Jeffries
> > <squid3 at treenet.co.nz <mailto:squid3 at treenet.co.nz>
> <mailto:squid3 at treenet.co.nz <mailto:squid3 at treenet.co.nz>>
> > > wrote:
> > >
> > > On 20/05/22 19:44, Praveen Ponakanti wrote:
> > > > Hi Alex,
> > > >
> > > > Thanks for going through several steps to help mitigate
> > src port
> > > > exhaustion. We are looking to achieve 400-500% more
> > > > concurrent connections if we could :) as there is a
> > > significant buffer
> > > > on the available CPU.
> > >
> > > Then you require at least 4, maybe 5, IP addresses to
> handle
> > that many
> > > concurrent connections with Squid.
> > >
> > >
> > > We would like to investigate going beyond the ephemeral port
> > range for
> > > some specific destination IP:PORT addresses. For that it
> appears
> > squid
> > > does not round-robin requests if we use multiple
> > tcp_outgoing_addresses.
> > > We could use ACL’s to pick a different outbound IP based
> on the
> > clients
> > > source IP, however that is not very ideal in our
> environment as our
> > > clients aren’t always equally split by subnet. However, if
> we could
> > > split by the client’s source port that might help achieve
> this. For
> > > example something like:
> > >
> > >
> > > acl pool1 clientport 0-32768
> > >
> > > acl pool2 clientport 32769-65536
> > >
> > >
> > > tcp_outgoing_address 10.1.0.1 pool1
> > >
> > > tcp_outgoing_address 10.1.0.2 pool2
> > >
> > >
> > > Squid's ACLs currently do not allow filtering by the
> client's source
> > > port. We could look into a separate patch to add this
> > functionality to
> > > squid’s ACL code if that makes sense. Or is there a better
> way to
> > > achieve this?
> > >
> > >
> > > Thanks
> > >
> > > Praveen
> > >
> > >
> > > > The option to use multiple tcp_outoing_addresses
> appears to be
> > > promising
> > > > along with some tweaks to the TCP timeouts. I guess we
> > could use
> > > ACLs to
> > > > pick a different outbound IP based on the
> requesting client's
> > > prefix. We
> > > > had not considered that option as the ephemeral
> ports were
> > no longer
> > > > available to other applications when squid uses most of
> > them with a
> > > > single outbound IP configured. We are also looking to
> > modify the
> > > code to
> > > > use the IP_BIND_ADDRESS_NO_PORT sockopt as that
> could help
> > delay
> > > port
> > > > assignment with the bind() call on the outbound TCP
> > sessions (to
> > > > hopefully allow access to the 4-tuple on the socket).
> > >
> > > Patches welcome.
> > >
> > > However, please be aware that use of the 4-tuple is
> often no
> > different
> > > from the 3-tuple since the dst-port is typically identical
> > for all
> > > outgoing traffic to a given dst-IP.
> > >
> > >
> > > Cheers
> > > Amos
More information about the squid-users
mailing list