Reducing overflowed and dropped work requests

[Václav Bílek]

increase the min thread per pool ( 100 or more) idle thread cost nothing
, spawning a new thread lasts...

Simon Males napsal(a):
> On Fri, Jul 15, 2011 at 11:15 AM, Simon Males <sime at sime.net.au> wrote:
>>> The Varnish munin plugins have alerted me a number of times (say less
>>> then 5) that N dropped work requests (n_wrk_drop) is greater then 1.
>>>
>>> According to Kristian Lyngstol's post [1] this is represents when the
>>> request queue it full. Additionally I believe the overflowed work
>>> requests (n_wrk_overflow) will always be greater then n_wrk_drop.
>>>
>>> I think I should be trying to tackle two issues: Limit requests
>>> overflowing, followed by increasing the size of the request queue.
>>>
>>> Of the below startup parameters I believe thread_pools should match
>>> the number of cores the server has and increase the maximum number of
>>> threads to possibly double (from 25 to 50?).
>>>
>>> DAEMON_OPTS="-a :80 \
>>>             -T localhost:6082 \
>>>             -f /etc/varnish/default.vcl \
>>>             -S /etc/varnish/secret \
>>>             -w 10,25,120 \
>>>             -p listen_depth=8192 \
>>>             -p log_hashstring=off \
>>>             -p lru_interval=60 \
>>>             -p sess_timeout=10 \
>>>             -p shm_workspace=32768 \
>>>             -p ping_interval=1 \
>>>             -p thread_pools=4 \
>>>             -s malloc,2G"
>> As overflow_max is default set to 100%, I believe I essentially need
>> to increase the number of worker threads to increase the ratio. To
>> increase the threads, I'll increase the thread_pools.
> 
> Since this post, I increased the thread_pools to 8, though today I've
> had another spurt of dropped work requests. It's definitely related to
> a sudden rise of traffic, as the dropped requests happen moment after
> sending an E-Mail to our users.
> 
> The servers have 16 cores, so I've increased the thread_pools to 16
> now. With 10 minimum threads per pool, I now have 160 threads waiting
> to serve.
> 
> Hopefully this will minimise the chance of dropping work requests.
>