Re: [PATCH RFC 1/2] qrwlock: A queue read/write lock implementation

From: Waiman Long
Date: Fri Jul 19 2013 - 11:30:27 EST

On 07/19/2013 04:40 AM, Ingo Molnar wrote:
* Waiman Long<waiman.long@xxxxxx> wrote:

On 07/18/2013 03:42 AM, Ingo Molnar wrote:
* Waiman Long<waiman.long@xxxxxx> wrote:

+ * stealing the lock if come at the right moment, the granting of the
+ * lock is mostly in FIFO order.
+ * 2. It is faster in high contention situation.
Again, why is it faster?
The current rwlock implementation suffers from a thundering herd
problem. When many readers are waiting for the lock hold by a writer,
they will all jump in more or less at the same time when the writer
releases the lock. That is not the case with qrwlock. It has been shown
in many cases that avoiding this thundering herd problem can lead to
better performance.
Btw., it's possible to further optimize this "writer releases the lock to
multiple readers spinning" thundering herd scenario in the classic
read_lock() case, without changing the queueing model.

Right now read_lock() fast path is a single atomic instruction. When a
writer releases the lock then it makes it available to all readers and
each reader will execute a LOCK DEC instruction which will succeed.

This is the relevant code in arch/x86/lib/rwlock.S [edited for
readability]:

__read_lock_failed():

0: LOCK_PREFIX
READ_LOCK_SIZE(inc) (%__lock_ptr)

1: rep; nop
READ_LOCK_SIZE(cmp) $1, (%__lock_ptr)
js 1b

LOCK_PREFIX READ_LOCK_SIZE(dec) (%__lock_ptr)
js 0b

ret

This is where we could optimize: instead of signalling to each reader that
it's fine to decrease the count and letting dozens of readers do that on
the same cache-line, which ping-pongs around the numa cross-connect
touching every other CPU as they execute the LOCK DEC instruction, we
could let the _writer_ modify the count on unlock in essence, to the exact
value that readers expect.

Since read_lock() can never abort this should be relatively
straightforward: the INC above could be left out, and the writer side
needs to detect that there are no other writers waiting and can set the
count to 'reader locked' value - which the readers will detect without
modifying the cache line:

__read_lock_failed():

0: rep; nop
READ_LOCK_SIZE(cmp) $1, (%__lock_ptr)
js 0b

ret

(Unless I'm missing something that is.)

That way the current write_unlock() followed by a 'thundering herd' of
__read_lock_failed() atomic accesses is transformed into an efficient
read-only broadcast of information with only a single update to the
cacheline: the writer-updated cacheline propagates in parallel to every
CPU and is cached there.

On typical hardware this will be broadcast to all CPUs as part of regular
MESI invalidation bus traffic.

reader unlock will still have to modify the cacheline, so rwlocks will
still have a fundamental scalability limit even in the read-only usecase.
I think that will work. The only drawback that I can see is the fairness
argument. The current read/write lock implementation is unfair to the
writer. That change will make it even more unfair to the writer and
there is no easy way to detect a waiting writer unless we change the
structure to add such a field. As a result, a steady stream of readers
will have a higher chance of blocking out a writer indefinitely.
The effect will have to be measured - but I don't think it's particularly
hard to tune the fairness balance between readers and writers: the change
I suggested would only affect the case when a writer already holding the
lock unlocks it.

But when a writer already holds the lock it can decide to pass that lock
to another writer-spinning instead of unlocking to all readers. This too
should be relatively straightforward to implement because neither
read_lock() nor write_lock() can abort and race.

Instead of doing:

static inline void arch_write_unlock(arch_rwlock_t *rw)
{
asm volatile(LOCK_PREFIX WRITE_LOCK_ADD(%1) "%0"
: "+m" (rw->write) : "i" (RW_LOCK_BIAS) : "memory");
}

the current owner could check whether there are other writers waiting and
could drop into a slowpath that passes ownership to one of the writers via
toggling bit 30 or so. This reduces the max number of writers by a factor
of 2.

But I'd implement this only if it proves to be a problem in practice.

I'd strongly suggest to first address the thundering herd problem of
write_unlock() and see how it affects scalability - before totally
replacing it all with a new, fundamentally heavier locking primitive!

Thanks,

Ingo
I had run some performance tests using the fserver and new_fserver benchmarks
(on ext4 filesystems) of the AIM7 test suite on a 80-core DL980 with
HT on. The following kernels were used:

1. Modified 3.10.1 kernel with mb_cache_spinlock in fs/mbcache.c
replaced by a rwlock
2. Modified 3.10.1 kernel + modified __read_lock_failed code as suggested
by Ingo
3. Modified 3.10.1 kernel + queue read/write lock
4. Modified 3.10.1 kernel + queue read/write lock in classic read/write
lock behavior

The last one is with the read lock stealing flag set in the qrwlock
structure to give priority to readers and behave more like the classic
read/write lock with less fairness.

The following table shows the averaged results in the 200-1000
user ranges:

+-----------------+--------+--------+--------+--------+
| Kernel | 1 | 2 | 3 | 4 |
+-----------------+--------+--------+--------+--------+
| fserver JPM | 245598 | 274457 | 403348 | 411941 |
| % change from 1 | 0% | +11.8% | +64.2% | +67.7% |
+-----------------+--------+--------+--------+--------+
| new-fserver JPM | 231549 | 269807 | 399093 | 399418 |
| % change from 1 | 0% | +16.5% | +72.4% | +72.5% |
+-----------------+--------+--------+--------+--------+

The following table shows the averaged results in the 1100-2000
user ranges:

+-----------------+--------+--------+--------+--------+
| Kernel | 1 | 2 | 3 | 4 |
+-----------------+--------+--------+--------+--------+
| fserver JPM | 230295 | 263562 | 347161 | 333908 |
| % change from 1 | 0% | +14.5% | +50.8% | +45.0% |
+-----------------+--------+--------+--------+--------+
| new-fserver JPM | 241154 | 274571 | 338166 | 369550 |
| % change from 1 | 0% | +13.9% | +40.2% | +53.2% |
+-----------------+--------+--------+--------+--------+

For these 2 benchmarks, the read/write lock bottleneck is in the
j_state_lock of the journal_s structure in include/linux/jbd2.h.

The following table show the amount of CPU time spent on the slowpaths
of the read and write lock for each of the different kernels listed
above as reported by perf with the fserver benchmark at 1500 users:

+-----------------+--------+--------+--------+--------+
| Kernel | 1 | 2 | 3 | 4 |
+-----------------+--------+--------+--------+--------+
| Read slowpath | 16.9% | 12.2% | 11.7% | 15.3% |
| Write slowpath | 13.3% | 11.5% | 0.02% | 0.09% |
+-----------------+--------+--------+--------+--------+

The small write slowpath numbers indicates that it is a reader heavy
lock with writers probably holding the lock a lot longer than the
readers.

It can be seen that the Ingo's changes does improve performance by
about 10-15% in this reader-heavy high contention scenario. It also reduces the read slowpath time relative to the write slowpath. However,
the queue read/write lock can improve performance by more than 50%. The
queue read/write lock is also much more fair to the writers as the
writers waiting time drop significantly.

In low contention situation, all the 4 kernels perform similarly with
1-2% performance variation and so it is hard to draw conclusion.

Regards,
Longman


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/