I've made some comments to the minor changes. See below. Still
looking at the big changes.
-sam
On Jul 18, 2006, at 4:00 PM, Julian Martin Kunkel wrote:
Hi,
enclosed you will find patches for the following issues:
Major changes:
* sync-coalesce:
If the last couple of operations which need to be synced finish
with an error
then the other operations can be stalled, too due to the handling
by the
request scheduler. (I have found this bug a couple of weeks ago but
did not
know what caused the race condition). It took me quite a while to
track that
bug...
The policy right now ensures that error requests never are enqueued.
(Note: should we flush the db in cases of an error ?)
*deleting in the background
Files are during the deletion renamed an deleted in the background
to shorten
response time of the server in case the file is very big...
Note: people have to put the storage dir into one filesystem.
* Trove multique support:
Now there is a thread for metadata ro, metadata rw, I/O and for
deleting in
the background.
This patch improves the throughput for read only ops, while write
ops happen.
Usually read ops are expected to be cached effectively, while all
write ops
force disk operations.
* Trove queues are changed to support an internal number of queued
elems.
Also I removed a few functions, added more prefixes with _nolock.
The make move_op_to_completion_queue is removed also some other
functions are
replaced with dbpf_move_op_to_completion_queue and
dbpf_op_pop_front_nolock.
* dbpf_dspace_cancel modified to guarantee that operation is not
finished
right now, also change to use id_safe_gen instead of fast_gen. This
changes
make the interface more useful and remove some dependencies on the
usage from
the upper layers. Also test, testsome changed that they can be
called on the
same time without possible memleaks / segfaults.
Minor changes:
* added define for mkdir syscall in dbpf.h
* added dbpf_op_get_status (and set status) to change status of
return value
atomically (this reduces the lines of code and makes the calls more
consistant).
*Stripped out non threaded Trove code.
Cool.
* Enhanced request scheduler debugging
Added a new function to pvfs2-internal.h server_op_to_str, which
allows to
fancy output the name of the op. The function is implemented in
PINT-reqproto-encode.c, maybe this is not the right place for the
function ?
This is used in the request scheduler, which now prints the whole
queue on
each enqueue op with current states of all ops for that particular
handle
(e.g. serviced or queued).
We've already got all the names of the operations in pvfs2-server.c
(see the PINT_server_req_table). I'd like to see that reused if
possible. Adding another array of all the operations means one more
place that will have to be modified when new operations are added.
* performance debug:
I want to add a performance debugging option which allows to print
a couple of
interesting metrics on the server side which might be analysed post
mortem.
(e.g. number of elements in the trove queues etc.)
For example one might run a skript to determine the time
distribution of sync
requests or I/O requests. I haven't added other logic yet but will,
soon.
I'd prefer to see this as part of the event handling code, instead of
as debug messages. If the event handling code doesn't have the
capabilities you need, we should fix it so it does.
* replaced fsync with fdatasync,
this might help on some journaling filesystems to improve
throughput because
metadata is not forced to be written.
Good idea. We should probably have a check for
_POSIX_SYNCHRONIZED_IO to make sure its there though.
-sam
enjoy,
Julian
<patchTrove.patch>
_______________________________________________
Pvfs2-developers mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
_______________________________________________
Pvfs2-developers mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
