glusterfs.git/xlators/performance/io-threads/src/io-threads.h, branch v3.1.0qa4

Memory accounting changes

2010-04-23T13:32:52+00:00

Memory accounting Changes. Thanks to Vinayak Hegde and Csaba Henk for their
contributions.

Signed-off-by: Vijay Bellur 
Signed-off-by: Anand V. Avati 

BUG: 329 (Replacing memory allocation functions with mem-type functions)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=329

io-threads: more fixes

2010-02-27T17:34:04+00:00

* conditional for scaling up threads was wrong

* ETIMEDOUT check was performed wrongly

Signed-off-by: Anand V. Avati 
Signed-off-by: Anand V. Avati 

BUG: 583 (filesystem access hangs while deleting large files)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=583

io-threads: single queue/multi-thread model

2010-02-22T16:51:40+00:00

This patch lets io-threads work with a single queue and multiple
threads work on picking the next request from the queue and process
it.

Whenever the number of pending requests in the queue double, a new
worker thread is spawned.

Workers expire after a (configurable) timeout of inactivity

Signed-off-by: Anand V. Avati 
Signed-off-by: Anand V. Avati 
Signed-off-by: Anand V. Avati 

BUG: 583 (filesystem access hangs while deleting large files)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=583

Changed occurrences of Z Research to Gluster.

2009-10-07T10:54:49+00:00

Signed-off-by: Anand V. Avati

io-threads: Fix missed notification in sem_timedwait

2009-06-11T13:29:17+00:00

We're performing a calculation for skewing idle time
that resulted in a timespec.tv_nsec value becoming larger than
1000 million or less than 0, forcing sem_timedwait to return
with an EINVAL instead of waiting for a request notification
from sem_post in iot_notify_worker(). This resulted in a missed
notification that resulted in a hang followed by a timeout
on the protocol/client side.

This commit avoids the over- and under-flow in tv_nsec by
skewing the tv_sec value instead.

Signed-off-by: Anand V. Avati

io-threads: Change mutexes/condvars to spinlocks/semaphores

2009-05-20T18:44:49+00:00

It seems that use of mutexes is resulting in pretty high thread
sleep and wake-up cost. What is worse, if a worker thread has
acquired a  lock, there is a possibility of the main glusterfs thread
being put to sleep. We change the use of mutexes into spinlock.
At the same time, we cannot anymore use condvars for notification since
the condvar interface depends on mutexes itself. Semaphores come to
out rescue. Luckily, even the pthread semaphores have a timedwait
interface to allow our idle worker threads to make an exit decision.

Further, it is possible that spinlocks are not available on all systems
so all this is curtained behind #defines so we can fall back to
mutexes and condvars implementation.

Signed-off-by: Anand V. Avati

io-threads: Reduce default max for threads to 64

2009-05-20T18:42:25+00:00

We've had complaints from users who've used
autoscaling option with default settings for min and
max threads, about high memory consumption because of
the large default value for max-threads.

Signed-off-by: Anand V. Avati

io-threads: Support mem-pool allocator for iot_request_t

2009-05-20T18:40:01+00:00

This commit brings in support for allocation of iot_request_t's
in io-threads through the use of the mem-pool. We're hoping
that the overheads of hundreds and thousands of small allocations
can be avoided through this.

The important point to note is that the memory pool is not
for the translator as a whole but there is one small memory
pool for each worker thread. Not only does that help us
avoid malloc overheads for small allocations like iot_request_t
but also avoid contention on the heap data structures when multiple
threads want an iot_request_t from the pool.

Signed-off-by: Anand V. Avati

io-threads: Add graceful shutdown of worker threads

2009-05-20T18:37:37+00:00

Signed-off-by: Anand V. Avati

io-threads: Clarify thread count range

2009-05-05T06:39:34+00:00

This patch cleans up io-threads behaviour regarding the
range values that can be specified for min-threads
and max-threads. THe major change is that the min threads
have been reduced to 2 to signify that io-threads needs minimum
two threads for its operation, while keeping the default number of
threads at 16. The idea is to decouple the default thread count
from the minimum thread count.

Note to Avati:
This applies over Raghu's indentation and logging take-3 patch.

Signed-off-by: Anand V. Avati