On 9/12/22 18:20, roee klinger wrote:
Thank you for your advice, as suggested I tested a both TCP and UDP, and
found TCP to be so slow
that is it practically unusable, UDP however works fine, but has the
normal disadvantages of UDP.
Glad that test yielded useful results in your environment!
It seems that TCP sends each message by itself, and UDP sends them in
bulk (10-20 at a time),
which seems to be a big part of the performance difference, in fact TCP
is so slow that it causes Squid to crash after a while.
After carefully reading the logs, I tried to get TCP to send in bulk as
well by configuring:
logformat logstash %ts.%03tu %6tr %>a %<la %>lp %Ss/%03>Hs %<st %rm
%ru %[un %Sh/%<a %mt %{Client-Tags}>h
access_log tcp://127.0.0.1:5400 logformat=logstash buffer-size=64KB
buffered_logs on
But it seems to still be sending the logs line by line, am I missing
something?
I have not tested this, but if my interpretation of the underlying code
is correct, then the TCP logging module code does not flush the buffer
when it accumulates (close to) buffer-size bytes. Instead, the module
starts writing its buffer after accumulating more than 16KB (2*MAX_URL).
How long are your typical access.log records?
Each write(2) call of a buffering TCP logging module during normal Squid
operation should be 16KB, regardless of the buffer-size setting.
FWIW, the official access_log buffer-size documentation hints at
module-specific flushing algorithms. I cannot describe the exact
algorithm used by the TCP logging module in a few words, but it is not
as simple as "accumulate buffer-size bytes and then write the
accumulated bytes".
HTH,
Alex.
On 9 Sep 2022, 19:57 +0300, Alex Rousskov wrote:
On 9/9/22 08:49, roee klinger wrote:
Hello,
I have just recently started exploring ingesting logs from Squid via
Logstash (TCP log ingestion)
I now have to make the decision of how to deploy Logstash, it seems to
me that I have two options:
1. Deploy a Logstash instance for every region where I have a Squid
instance running, resulting in the lowest latency possible between
Squid and Logstash.
2. Deploy a central Logstash instance for all Squid instance worldwide,
which is the cheapest and easiest option, but will have high latency
for some of the Squid instance which are located far away
geographically from the Logstash instance.
I already know that if the log server crashes in Squid, then Squid will
crash with a fatal error,
FYI: Logging error handling is configurable via "access_log on-error".
so Squid takes logging very seriously, so my
question is: does TCP logging performance in Squid effect the total
request performance, or are they independent from each other?
TCP logger uses asynchronous non-blocking I/O. Individual HTTP
transactions do not wait for individual log records to be sent. Needless
to say, logging does consume resources and, hence, may affect overall
HTTP transaction performance. And if the log record recipient is too
slow, then Squid will die and/or log records will be dropped.
What would be you recommendations, if request performance is crucial
(and log performance is not)?
I would not be surprised if you would not be able to measure meaningful
performance difference among these modules. From general performance
point of view, and module-specific bugs and limitations notwithstanding,
I would expect "daemon" and "udp" modules to have smaller performance
overhead/cost for Squid workers than the "tcp" module because I expect
that TCP has to "do more" than stdout I/O and UDP. However, I would not
base this decision on such high-level speculations and test various
options instead.
HTH,
Alex.
_______________________________________________
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
