On 15/03/11 00:02, Saurabh Agarwal wrote:
Hi All
I am trying to load test squid using this simple test. From a single
client machine I want to simultaneously download 200 different HTTP
files of 10MB each in a loop over and over again. I see that within 5
minutes squid process size goes beyond 250MB. These 10MB files are
all cachable and return a TCP_HIT for the second time onwards. There
are other processes running and I want to limit squid memory usage to
120MB. Hard disk partition allocated to Squid is of 10GB and is made
using device-mapper. I am using 3 cache_dir as mentioned below. How
can I control Squid memory usage in this case? Below is my portion of
my squid.conf.
200 files @ 10MB -> up to 2GB of data possibly in memory simultaneously.
It is easy to see why "squid process size goes beyond 250MB" easily.
You have cache_mem of 8 MB. Which means Squid will push these objects to
disk after the first use. From then on what you are testing is the rate
at which Squid can load them from disk into the network. It is quite
literally a read from disk into buffer, call function which immediately
writes direct from buffer to network. Done ins "small" chunks of
whatever the system disk I/O page size is (default 4KB but could be more).
The real speed bottleneck in Squid are the HTTP processing. Which does
a lot of CPU intensive small steps of parsing and data copying. When
there are a lot of new requests arriving it sucks CPU time away from
that speedy read->write byte pumping loop.
Your test is a classic check for Disk speed limits in Squid.
The other tests you need to check performance are:
* numerous requests for few medium sized objects (which can all fit in
memory together, headers of ~10% or less the total object). Testing the
best-case memory-hit speed.
* numerous requests for very small objects (one packet responses sort
of size). Testing the worst-case HTTP parser limits.
* parallel requests for numerous varied objects (too many to fit in
memory). Testing a somewhat normal traffic speed expectations.
There is a tool called WebPolygraph which does some good traffic
measurements.
---------------------------------------- access_log
/squid/logs/access.log squid cache_log /squid/logs/cache.log
cache_mem 8 MB cache_dir aufs /squid/var/cache/small 1500 9 256
max-size=10000 cache_dir aufs /squid/var/cache/medium 2500 6 256
max-size=20000000 cache_dir aufs /squid/var/cache/large 6000 3 256
max-size=100000000 maximum_object_size 100 MB log_mime_hdrs off
max_open_disk_fds 400 maximum_object_size_in_memory 8 KB
cache_store_log none pid_filename /squid/logs/squid.pid debug_options
ALL,1 -------------------
Regards, Saurabh
Um, your use of cache_dir is a bit odd.
*one* ufs/aufs/diskd cache_dir entry per disk spindle. Otherwise your
speed is lower due to disk I/O collisions between the cache_dir (your
test objects are all the same size and so will not reveal this behaviour).
Also, leave some disk space for the cache log and journal overheads.
Otherwise your Squid will crash with "unable to write to file" errors
when the cache starts to get nearly full.
Amos
--
Please be using
Current Stable Squid 2.7.STABLE9 or 3.1.11
Beta testers wanted for 3.2.0.5
