| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
| |
Going by the memory usage for each threads, it is prudent to
have lower number of threads by default and let users who understand
the memory consequences increase the thread count for themselves.
Signed-off-by: Anand V. Avati <avati@amp.gluster.com>
|
| |
|
|
|
|
|
|
| |
The default stack size on Linux is around 8 MiB for each
thread. This is clearly too high for our purpose. This commit reduces
the stack size down to 1 MiB.
Signed-off-by: Anand V. Avati <avati@amp.gluster.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This commit finally makes the autoscaling feature visible to the user.
Know that we're now using two separate thread-pools, one for data
requests, called ordered thread-pool in io-threads, and the other
for meta-data requests, called un-ordered thread-pool.
We do not expose this information to the user to keep io-threads
simple. Consequently, when the user specifies a min-threads and
max-threads value, the number of threads assigned to each pool
is equal, i.e. both pools start with their min threads set to half of
the option "min-threads" and both scale up their threads at most up to
half of option "max-threads".
Volfile options will be added to the wiki and user-guide.
Signed-off-by: Anand V. Avati <avati@amp.gluster.com>
|
| |
|
|
|
|
|
|
| |
The default is also to provide no scaling. For both, ordered and
unordered request pools, when scaling is off, we maintain atleast the
minimum number of threads specified in the volfile.
Signed-off-by: Anand V. Avati <avati@amp.gluster.com>
|
| |
|
|
|
|
|
|
|
|
|
| |
Now we have the remaining fops going through the ordered
thread-pool.
To route a request through ordered thread, we use
iot_schedule_ordered(..) and the worker thread for
ordered requests is iot_worker_ordered(..)
Signed-off-by: Anand V. Avati <avati@amp.gluster.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This commit adds everything needed to:
a. Get un-ordered request going through the un-ordered
thread-pool. This happens through, the
iot_schedule_unordered(..). The unordered thread-pool
consists of thread running the iot_worker_unordered(..)
function.
b. Make threads in the un-ordered thread pool start-up
and exit depending on the thread state.
Note that at this point the requests that need
ordering are still going through iot_schedule(..).
Signed-off-by: Anand V. Avati <avati@amp.gluster.com>
|
| |
|
|
| |
Signed-off-by: Anand V. Avati <avati@amp.gluster.com>
|
| |
|
|
|
|
|
|
|
|
| |
Worker threads were represented as a list in iot_conf_t
which made us traverse the list of workers in order to
decide which thread gets the request. Now we represent the
workers as a dynamically allocated array so that we can just index
into the array to schedule the file.
Signed-off-by: Anand V. Avati <avati@amp.gluster.com>
|
| |
|
|
|
|
|
|
| |
This patch changes the per-thread request queue from a custom circular
linked list, into the standard list.h list which is easier to
understand and has a cleaner interface.
Signed-off-by: Anand V. Avati <avati@amp.gluster.com>
|
| |
|
|
|
|
| |
updated copyright header to include 2009.
Signed-off-by: Anand V. Avati <avati@amp.gluster.com>
|
| |
|
|
|
|
|
|
|
|
|
| |
iot_queue() and iot_dequeue() functions were using a io-threads
translator-wide lock which would be contended for by every worker
thread waiting for IO requests.
This patch reduces the granularity by turning the
lock into a per-worker lock.
Signed-off-by: Anand V. Avati <avati@dev.gluster.com>
|
| |
|
|
|
|
|
|
| |
Since we're not dependent on this io-thread internal state(i.e.
cache_size and current_size) to rate limit requests, we can remove
these two data members and code that checks for these.
Signed-off-by: Anand V. Avati <avati@dev.gluster.com>
|
| |
|
|
| |
Signed-off-by: Anand V. Avati <avati@dev.gluster.com>
|
| |
|
