On Tue, Jul 03, 2012 at 11:59:51AM +0100, Mel Gorman wrote:
On Tue, Jul 03, 2012 at 10:19:28AM +1000, Dave Chinner wrote:
On Mon, Jul 02, 2012 at 08:35:16PM +0100, Mel Gorman wrote:
Adding dri-devel and a few others because an i915 patch contributed to the regression.
On Mon, Jul 02, 2012 at 03:32:15PM +0100, Mel Gorman wrote:
On Mon, Jul 02, 2012 at 02:32:26AM -0400, Christoph Hellwig wrote:
It increases the CPU overhead (dirty_inode can be called up to 4 times per write(2) call, IIRC), so with limited numbers of threads/limited CPU power it will result in lower performance. Where you have lots of CPU power, there will be little difference in performance...
When I checked it it could only be called twice, and we'd already optimize away the second call. I'd defintively like to track down where the performance changes happend, at least to a major version but even better to a -rc or git commit.
By all means feel free to run the test yourself and run the bisection :)
It's rare but on this occasion the test machine is idle so I started an automated git bisection. As you know the milage with an automated bisect varies so it may or may not find the right commit. Test machine is sandy so http://www.csn.ul.ie/~mel/postings/mmtests-20120424/global-dhp__io-metadata-... is the report of interest. The script is doing a full search between v3.3 and v3.4 for a point where average files/sec for fsmark-single drops below 25000. I did not limit the search to fs/xfs on the off-chance that it is an apparently unrelated patch that caused the problem.
It was obvious very quickly that there were two distinct regression so I ran two bisections. One led to a XFS and the other led to an i915 patch that enables RC6 to reduce power usage.
[aa464191: drm/i915: enable plain RC6 on Sandy Bridge by default]
Doesn't seem to be the major cause of the regression. By itself, it has impact, but the majority comes from the XFS change...
The fact it has an impact at all is weird but lets see what the DRI folks think about it.
[c999a223: xfs: introduce an allocation workqueue]
Which indicates that there is workqueue scheduling issues, I think. The same amount of work is being done, but half of it is being pushed off into a workqueue to avoid stack overflow issues (*). I tested the above patch in anger on an 8p machine, similar to the machine you saw no regressions on, but the workload didn't drive it to being completely CPU bound (only about 90%) so the allocation work was probably always scheduled quickly.
What test were you using?
fsmark, dbench, compilebench, and a few fio workloads. Also, xfstests times each test and I keep track of overall runtime, and none of those showed any performance differential, either...
Indeed, running on a current 3.5-rc5 tree, my usual fsmark benchmarks are running at the same numbers I've been seeing since about 3.0 - somewhere around 18k files/s for a single thread, and 110-115k files/s for 8 threads.
I just ran your variant, and I'm getting about 20kfile/s for a single thread, which is about right because you're using smaller directories than I am (22500 files per dir vs 100k in my tests).
How many worker threads have been spawned on these machines that are showing the regression?
20 or 21 generally. An example list as spotted by top looks like
Pretty normal.
What is the context switch rate on the machines whenteh test is running?
.....
Vanilla average context switch rate 4278.53 Revert average context switch rate 1095
That seems about right, too.
Can you run latencytop to see if there is excessive starvation/wait times for allocation completion?
I'm not sure what format you are looking for.
Where the context switches are coming from, and how long they are abeing stalled for. Just to get the context switch locations, you can use perf on the sched:sched_switch event, but that doesn't give you stall times. Local testing tells me that about 40% of the switches are from xfs_alloc_vextent, 55% are from the work threads, and the rest are CPU idling events, which is exactly as I'd expect.
A pert top profile comparison might be informative, too...
I'm not sure if this is what you really wanted. I thought an oprofile or perf report would have made more sense but I recorded perf top over time anyway and it's at the end of the mail.
perf report and oprofile give you CPU usage across the run, it's not instantaneous and that's where all the interesting information is. e.g. a 5% sample in a 20s profile might be 5% per second for 20s, or it might be 100% for 1s - that's the behaviour run profiles cannot give you insight into....
As it is, the output you posted is nothing unusual.
For just these XFS tests I've uploaded a tarball of the logs to http://www.csn.ul.ie/~mel/postings/xfsbisect-20120703/xfsbisect-logs.tar.gz
Ok, so the main thing I missed when first looking at this is that you are concerned about single thread regressions. Well, I can't reproduce your results here. Single threaded with or without the workqueue based allocation gives me roughly 20k +/-0.5k files/s one a single disk, a 12 disk RAID0 array and a RAM disk on a 8p/4GB RAM machine. That's the same results I've been seeing since I wrote this patch almost 12 months ago....
So, given that this is a metadata intensive workload, the only extent allocation is going to be through inode and directory block allocation. These paths do not consume a large amount of stack, so we can tell the allocator not to switch to workqueue stack for these allocations easily.
The patch below does this. It completely removes all the allocation based context switches from the no-data fsmark workloads being used for this testing. It makes no noticable difference to performance here, so I'm interested if it solves the regression you are seeing on your machines.
Cheers,
Dave.