Andrei, regarding core binding, our test program has only 1 thread
which appends to the log. I did not explicitly bind this test to any
core. Even without core binding, would a thread already scheduled for
the core (say due to NUMA) get stuck until the core is available or
does the kernel migrate this poor thread?

Regarding journaling, your explanation is very clear. I read that a
record takes up 1 to multiple buffers, 4 KB each. And a journal
default size is 128 MB, regardless of partition size. Thus a journal
can be filled up to 32,000  records before it has to check point. Is
this correct?

Thanks again for your help.

--Cuong

On Thu, Sep 12, 2013 at 8:47 AM, Sidorov, Andrei
[off-list ref] wrote:

If there are threads bound to this core, they probably won't be able to
progress for the time of commit. I don't think scheduler would migrate
them to a different core immediately. So, if GC wants to talk to these
threads, it would block as well.

Of course deleting many small files is much slower than deleting single
file of the same size. Number of records is roughly the number of files
deleted, split at allocation group boundaries. The group size is 128M
unless you enabled bigalloc feature. That is in simplest case a release
blocks entry for 256M file contains one record with 2 groups to free.
But release blocks entry for 256 1M files will contain 256 records that
are to be precessed separately. Before that change, commit time for
removing one 256M is the same as removing 256 1M files (wrt to time
taken to release blocks). After the change, releasing blocks of 256M
file would take 2 “iterations” as opposed to 256.

Basically, there is no thing like “deleting N blocks”, unless you delete
a file by progressively truncating it towards zero.

Regards,
Andrei.

On 12.09.2013 02:08, Cuong Tran wrote:

quoted

My desk top has 8 cores, including hyperthreading. Thus deleting files
would lock up one core but that should not affect GC threads if core
lock-up is an issue? Would # journal records be proportional to #
blocks deleted. And thus deleting N blocks, one block at a time would
create N times more journal records than deleting all N blocks in "one
shot"?

--Cuong

On Wed, Sep 11, 2013 at 11:02 PM, Sidorov, Andrei
[off-list ref] wrote:

quoted

It would lock-up one core whichever jdb/sdaX runs on. This will usually
happen upon commit that runs every x seconds, 5 by default (see “commit”
mount option for ext4). I.e. deleting 5 files one by one with 1 second
interval in between is basically the same as deleting all of them “at once”.

Yes, fallocated files are the same wrt releasing blocks.

Regards,
Andrei.

On 12.09.2013 01:45, Cuong Tran wrote:

quoted

Awesome fix and thanks for very speedy response. I have some
questions. We delete files one at a time, and thus that would lock up
one core or all cores?

And in our test, we use falloc w/o writing to file. That would still
cause freeing block-by-block, correct?
--Cuong

On Wed, Sep 11, 2013 at 10:32 PM, Sidorov, Andrei
[off-list ref] wrote:

quoted

Hi,

Large file deletions are likely to lock cpu for seconds if you're
running non-preemptible kernel < 3.10.
Make sure you have this change:
http://patchwork.ozlabs.org/patch/232172/ (available in 3.10 if I
remember it right).
Turning on preemption may be a good idea as well.

Regards,
Andrei.

On 12.09.2013 00:18, Cuong Tran wrote:

quoted

We have seen GC stalls that are NOT due to memory usage of applications.

GC log reports the CPU user and system time of GC threads, which are
almost 0, and stop-the-world time, which can be multiple seconds. This
indicates GC threads are waiting for IO but GC threads should be
CPU-bound in user mode.

We could reproduce the problems using a simple Java program that just
appends to a log file via log4j. If the test just runs by itself, it
does not incur any GC stalls. However, if we run a script that enters
a loop to create multiple large file via falloc() and then deletes
them, then GC stall of 1+ seconds can happen fairly predictably.

We can also reproduce the problem by periodically switch the log and
gzip the older log. IO device, a single disk drive, is overloaded by
FS flush when this happens.

Our guess is GC has to acquiesce its threads and if one of the threads
is stuck in the kernel (say in non-interruptible mode). Then GC has to
wait until this thread unblocks. In the mean time, it already stops
the world.

Another test that shows similar problem is doing deferred writes to
append a file. Latency of deferred writes is very fast but once a
while, it can last more than 1 second.

We would really appreciate if you could shed some light on possible
causes? (Threads blocked because of journal check point, delayed
allocation can't proceed?). We could alleviate the problem by
configuring expire_centisecs and writeback_centisecs to flush more
frequently, and thus even-out the workload to the disk drive. But we
would like to know if there  is a methodology to model the rate of
flush vs. rate of changes and IO throughput of the drive (SAS, 15K
RPM).

Many thanks.
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html