Probably all, but observed on Ubuntu 12.04, OpenJDK 6

Sorry, I typo'd the example. (defn total ...) should be (defn sum ...).

fixed typeo in example

Couldn't reproduce the exception, but the 2nd example did chew through about 4x the amount of memory. Vetting.

adding a patch. Since most of Clojure ends up running this code in one way or another, I'd assert that tests are included as part of the normal Clojure test process.

Patch simply calls Util.ret1(argx, argx=null) on all invoke calls.

And as a note, both examples in the original report now have extremely similar memory usages.

Sounds great, and the patch looks good too. Let me know if I need to do anything else.

clj-1087-diff-perf-enhance-patch-v1.txt dated Oct 15 2012 implements Tom's suggested performance enhancement, although not exactly in the way he suggested. It does calculate the union of the two key sequences.

Please close and reject. The patch is not working if val has non final fields.

Java(TM) SE Runtime Environment (build 1.7.0_04-b21)
Java HotSpot(TM) 64-Bit Server VM (build 23.0-b21, mixed mode)

I added a patch consisting of three commits:

• one which adds caching to seqs, sets, maps, vectors and queues
• one that aligns the shape of Util/hasheq on the one Util/equiv (to have a consistent behavior from the JIT compiler: without that deoptimization was more penalizing for hasheq than for equiv)
• one that fix hasheq on records (which was non consistent with its equiv impl.) – and this commit relies on a static method introduced in the "caching hasheq" commit

In the process of screening this, I'm not seeing much of a performance difference after applying the patch.

before patch:
user=> (-main)

Version: 1.5.0-master-SNAPSHOT
"Elapsed time: 6373.752 msecs"
"Elapsed time: 6578.037 msecs"
"Elapsed time: 6476.399 msecs"

after patch:
user=> (-main)

Version: 1.5.0-master-SNAPSHOT
"Elapsed time: 6182.699 msecs"
"Elapsed time: 6548.086 msecs"
"Elapsed time: 6496.711 msecs"

clojure 1.4:
user=> (-main)

Version: 1.4.0
"Elapsed time: 6484.234 msecs"
"Elapsed time: 6243.672 msecs"
"Elapsed time: 6248.898 msecs"

clojure 1.3
user=> (-main)

Version: 1.3.0
"Elapsed time: 3584.966 msecs"
"Elapsed time: 3618.189 msecs"
"Elapsed time: 3372.979 msecs"

I blew away my local clojure repo and re-applied the patch just to make sure, but the results are the same. Does this fix not optimize the case given in the original test project?

For reference I'm running this code:

(defn -main
[& args]

(println)
(println "Version: " (clojure-version))

(def mm 10000)

(def str-keys (map str (range mm)))
(def m (zipmap str-keys (range mm)))
(time (dotimes [i mm] (doseq [k str-keys] (m k))))

(def kw-keys (map #(keyword (str %)) (range mm)))
(def m (zipmap kw-keys (range mm)))
(time (dotimes [i mm] (doseq [k kw-keys] (m k))))

(def sym-keys (map #(symbol (str %)) (range mm)))
(def m (zipmap sym-keys (range mm)))
(time (dotimes [i mm] (doseq [k sym-keys] (m k))))

(println))

Sorry, I was too quick to react on the ML (someone said it was related to hasheq caching and since I had the patch almost ready: on a project I noticed too much time spent computing hasheq on vectors).
So no, my patch doesn't improve anything with kws, syms or strs as keys. However when keys are collections, it fares better.

In 1.4, for a "regular" object, it must fails two instanceof tests before calling .hashCode().
If we make keywords and symbols implement IHashEq and reverse the test order (first IHashEq, then Number) it should improve the performance of the above benchmark – except for Strings.

Marking as incomplete, should we also delete the patch as it seems like it should be in a different ticket?

In 1.3, #'hash was going through Object.hashCode and thus was simple and fast. Plus collections hashes were cached.
In 1.4 and master, #'hash goes now through two instanceof test (Number and IHasheq in this order) before trying Object.hashCode in last resort. Plus collections hashes are not cached.
As such I'm not sure Util.hasheq inherent slowness (compared to Util.hash) and hasheq caching should be separated in two issues.

The caching-hasheq-v2.diff patchset reintroduces hashes caching for collections/hasheq and reorders the instanceof tests (to test for IHashEq before Number) and makes Keyword and Symbol implement IHashEq to branch fast in Util.hasheq.

I recommend adding a collection test to the current benchmark:

(defn -main
[& args]

(println)
(println "Version: " (clojure-version))

(def mm 10000)

(def str-keys (map str (range mm)))
(def m (zipmap str-keys (range mm)))
(time (dotimes [i mm] (doseq [k str-keys] (m k))))

(def kw-keys (map #(keyword (str %)) (range mm)))
(def m (zipmap kw-keys (range mm)))
(time (dotimes [i mm] (doseq [k kw-keys] (m k))))

(def sym-keys (map #(symbol (str %)) (range mm)))
(def m (zipmap sym-keys (range mm)))
(time (dotimes [i mm] (doseq [k sym-keys] (m k))))

(def vec-keys (map (comp (juxt keyword symbol identity) str) (range mm)))
(def m (zipmap vec-keys (range mm)))
(time (dotimes [i mm] (doseq [k vec-keys] (m k))))

(println))

For some reason I can't get v2 to build against master. It applies cleanly, but fails to build.

Timothy: I inadvertently deleted a "public" modifier before commiting... fixed in caching-hasheq-v3.diff

I now get the following results:

Version: 1.4.0
"Elapsed time: 6281.345 msecs"
"Elapsed time: 6344.321 msecs"
"Elapsed time: 6108.55 msecs"
"Elapsed time: 36172.135 msecs"

Version: 1.5.0-master-SNAPSHOT (pre-patch)
"Elapsed time: 6126.337 msecs"
"Elapsed time: 6320.857 msecs"
"Elapsed time: 6237.251 msecs"
"Elapsed time: 18167.05 msecs"

Version: 1.5.0-master-SNAPSHOT (post-patch)
"Elapsed time: 6501.929 msecs"
"Elapsed time: 3861.987 msecs"
"Elapsed time: 3871.557 msecs"
"Elapsed time: 5049.067 msecs"

Marking as screened

Please, could you add as a comment the bench result using 1.3 vs 1.5-master-post-patch?

The performance with 1.5-master is now very close to 1.3 for 3/4 of the benchmark.

However, this code is still showing 43% performance drop (3411 ms VS 6030 ms – 1.3 VS 1.5-master):

(def str-keys (map str (range mm)))
(def m (zipmap str-keys (range mm)))
(time (dotimes [i mm] (doseq [k str-keys] (m k))))

Version: 1.3.0
"Elapsed time: 3411.353 msecs" <---
"Elapsed time: 3459.992 msecs"
"Elapsed time: 3365.182 msecs"
"Elapsed time: 3813.637 msecs"

Version: 1.4.0
"Elapsed time: 5710.073 msecs" <---
"Elapsed time: 5817.356 msecs"
"Elapsed time: 5774.856 msecs"
"Elapsed time: 18754.482 msecs"

Version: 1.5.0-master-SNAPSHOT
"Elapsed time: 6030.247 msecs" <---
"Elapsed time: 3372.637 msecs"
"Elapsed time: 3267.481 msecs"
"Elapsed time: 3852.927 msecs"

To reproduce:
$ git clone https://github.com/clojure/clojure.git
$ cd clojure
$ mvn install -Dmaven.test.skip=true

$ cd ..

$ git clone https://github.com/oshyshko/clj-perf.git
$ cd clj-perf
$ lein run-all

http://groups.google.com/group/clojure-dev/browse_thread/thread/6a8219ae3d4cd0ae?hl=en

Here's a stab in the dark attempt at rewriting MultiFn to use atoms to swap between immutable copies of its otherwise mutable state.

The four pieces of the MultiFn's state that are mutable and protected by locks are now contained in the MultiFn.State class, which is immutable and contains convenience functions for creating a new one with one field changed. An instance of this class is now held in an atom in the MultiFn called state. Changes to any of these four members are now done with an atomic swap of these State objects.

The getMethod/findAndCacheBestMethod complex was rewritten to better enable the atomic logic. findAndCacheBestMethod was replaced with findBestMethod, which merely looks up the method; the caching logic was moved to getMethod so that it can be retried easily as part of the work that method does.

As it was findAndCacheBestMethod seemingly had the potential to cause a stack overflow in a perfect storm of heavy concurrent modification, since it calls itself recursively if it finds that the hierarchy has changed while it has done its work. This logic is now done in the CAS looping of getMethod, so hopefully that is not even an unlikely possibility anymore.

There is still seemingly a small race here, since the check is done of a regular variable against the value in a ref. Now as before, the ref could be updated just after you do the check, but before the MultiFn's state is updated. Of course, only the method lookup part of a MultiFn call was synchronized before; it could already change after the lookup but before the method itself executed, having a stale method finish seemingly after the method had been changed. Things are no different now in general, with the atom-based approach, so perhaps this race is not a big deal, as a stale value can't persist for long.

The patch passes the tests and Clojure and multimethods seems to work.

this patch gets rid of the ugly lock contention issues. I have not been able to benchmark it vs. the old multimethod implementation, but looking at it, I would not be surprised if it is faster when the system is in a steady state.

This looks straightforward, except for getMethod. Can somebody with context add more discussion of how method caching works?

Also, it would be great to have tests with this one.

Obviously I didn't write the original code, so I'm not the ideal
person to explain this stuff. But I did work with it a bit recently,
so in the hopes that I can be helpful, I'm writing down my
understanding of the code as I worked with it. Since I came to the
code and sort of reverse engineered my way to this understanding,
hopefully laying this all out will make any mistakes or
misunderstandings I may have made easier to catch and correct. To
ensure some stability, I'll talk about the original MultiFn code as it
stands at this commit:
https://github.com/clojure/clojure/blob/8fda34e4c77cac079b711da59d5fe49b74605553/src/jvm/clojure/lang/MultiFn.java

There are four pieces of state that change as the multimethod is either
populated with methods or called.

• methodTable: A persistent map from a dispatch value (Object) to
  a function (IFn). This is the obvious thing you think it is,
  determining which dispatch values call which function.
• preferTable: A persistent map from a dispatch value (Object) to
  another value (Object), where the key "is preferred" to the value.
• methodCache: A persistent map from a dispatch value (Object) to
  function (IFn). By default, the methodCache is assigned the same
  map value as the methodTable. If values are calculated out of the
  hierarchy during a dispatch, the originating value and the
  ultimate method it resolves to are inserted as additional items in
  the methodCache so that subsequent calls can jump right to the
  method without recursing down the hierarchy and preference table.
• cachedHierarchy: An Object that refers to the hierarchy that is
  reflected in the latest cached values. It is used to check if the
  hierarchy has been updated since we last updated the cache. If it
  has been updated, then the cache is flushed.

I think reset(), addMethod(), removeMethod(), preferMethod(),
prefers(), isA(), dominates(), and resetCache() are extremely
straightforward in both the original code and the patch. In the
original code, the first four of those are synchronized, and the other
four are only called from methods that are synchronized (or from
methods only called from methods that are synchronized).

Invoking a multimethod through its invoke() function will call
getFn(). getFn() will call the getMethod() function, which is
synchronized. This means any call of the multimethod will wait for and
take a lock as part of method invocation. The goal of the patch in
this issue is to remove this lock on calls into the multimethod. It in
fact removes the locks on all operations, and instead keeps its
internal mutable state by atomically swapping a private immutable
State object held in an Atom called state.

The biggest change in the patch is to the
getFn()>getMethod()>findAndCacheBestMethod() complex from the
original code. I'll describe how that code works first.

In the original code, getFn() does nothing but bounce through
getMethod(). getMethod() tries three times to find a method to call,
after checking that the cache is up to date and flushing it if it
isn't:

1. It checks if there's a method for the dispatch value in the
methodCache.

2. If not, it calls findAndCacheBestMethod() on the
dispatch value. findAndCacheBestMethod() does the following:

1. It iterates through every map entry in the method table,
keeping at each step the best entry it has found so far
(according to the hierarchy and preference tables).

2. If it did not find a suitable entry, it returns null.

3. Otherwise, it checks if the hierarchy has been changed since the cache
was last updated. If it has not changed, it inserts the method
into the cache and returns it. If it has been changed, it
resets the cache and calls itself recursively to repeat the process.

3. Failing all else, it will return the method for the default
dispatch value.

Again, remember everything in the list above happens in a call to a
synchronized function. Also note that as it is currently written,
findAndCacheBestMethod() uses recursion for iteration in a way that
grows the stack. This seems unlikely to cause a stack overflow unless
the multimethod is getting its hierarchy updated very rapidly for a
sustained period while someone else tries to call it. Nonetheless, the
hierarchy is held in an IRef that is updated independently of the
locking of the MultiFn. Finally, note that the multimethod is locked
only while the method is being found. Once it is found, the lock is
released and the method actually gets called afterwards without any
synchronization, meaning that by the time the method actually
executes, the multimethod may have already been modified in a way that
suggests a different method should have been called. Presumably this
is intended, understood, and not a big deal.

Moving on now to the patch in this issue. As mentioned, the main
change is updating this entire apparatus to work with a single atomic
swap to control concurrency. This means that all updates to the
multimethod's state have to happen at one instant in time. Where the
original code could make multiple changes to the state at different
times, knowing it was safely protected by an exclusive lock, rewriting
for atom swaps requires us to reorganize the code so that all updates
to the state happen at the same time with a single CAS.

To implement this change, I pulled the implicit looping logic from
findAndCacheBestMethod() up into getMethod() itself, and broke the
"findBestMethod" part into its own function, findBestMethod(), which
makes no update to any state while implementing the same
logic. getMethod() now has an explicit loop to avoid stack-consuming
recursion on retries. This infinite for loop covers all of the logic
in getMethod() and retries until a method is successfully found and a
CAS operation succeeds, or we determine that the method cannot be
found and we return the default dispatch value's implementation.

I'll now describe the operation of the logic in the for loop. The
first two steps in the loop correspond to things getMethod() does
"before" its looping construct in the original code, but we have to do
in the loop to get the latest values.

1. First we dereference our state, and assign this value to both
oldState and newState. We also set a variable called needWrite to
false; this is so we can avoid doing a CAS (they're not free) when
we have not actually updated the state.

2. Check if the cache is stale, and flush it if so. If the cache
gets flushed, set needWrite to true, as the state has changed.

3. Check if the methodCache has an entry for this dispatch
value. If so, we are "finished" in the sense that we found the
value we wanted. However, we may need to update the state. So,

• If needWrite is false, we can return without a CAS, so just
  break out of the loop and return the method.

• Otherwise, we need to update the state object with a CAS. If
  the CAS is successful, break out of the loop and return the
  target function. Otherwise, continue on the next iteration
  of the loop, skipping any other attempts to fetch the method
  later in the loop (useless work, at this point).

4. The value was not in the methodCache, so call the function
findBestMethod() to attempt to find a suitable method based on the
hierarchy and preferences. If it does find us a suitable method,
we now need to cache it ourselves. We create a new state object
with the new method cache and attempt to update the state atom
with a CAS (we definitely need a write here, so no need to check
needWrite at this point).

The one thing that is possibly questionable is the check at this
point to make sure the hierarchy has not been updated since the
beginning of this method. I inserted this here to match the
identical check at the corresponding point in
findAndCacheBestMethod() in the original code. That is also a
second check, since the cache is originally checked for freshness
at the very beginning of getMethod() in the original code. That
initial check happens at the beginning of the loop in the
patch. Given that there is no synchronization with the method
hierarchy, it is not clear to me that this second check is needed,
since we are already proceeding with a snapshot from the beginning
of the loop. Nonetheless, it can't hurt as far as I can tell, it
is how the original code worked, and I assume there was some
reason for that, so I kept the second check.

5. Finally, if findBestMethod() failed to find us a method for the
dispatch value, find the method for the default dispatch value and
return that by breaking out of the loop.

So the organization of getMethod() in the patch is complicated by two
factors: (1) making the retry looping explicit and stackless, (2)
skipping the CAS when we don't need to update state, and (3) skipping
needless work later in the retry loop if we find a value but are
unable to succeed in our attempt to CAS. Invoking a multimethod that
has a stable hierarchy and a populated cache should not even have a
CAS operation (or memory allocation) on this code path, just a cache
lookup after the dispatch value is calculated.

I've updated this patch (removing the old version, which is entirely superseded by this one). The actual changes to MultiFn.java are identical (modulo any thing that came through in the rebase), but this newer patch has tests of basic multimethod usage, including defmulti, defmethod, remove-method, prefer-method and usage of these in a hierarchy that updates in a way interleaved with calls.

I created a really, really simple benchmark to make sure this was an improvement. The following tests were on a quad-core hyper-threaded 2012 MBP.

With two threads contending for a simple multimethod:
The lock-based multifns run at an average of 606ms, with about 12% user, 15% system CPU at around 150%.
The lockless multifns run at an average of 159ms, with about 25% user, 3% system CPU at around 195%.

With four threads contending for a simple multimethod:
The lock-based multifns run at an average of 1.2s, with about 12% user, 15% system, CPU at around 150%.
The lockless multifns run at an average of 219ms, with about 50% user, 4% system, CPU at around 330%.

You can get the code at https://github.com/davidsantiago/multifn-perf

It's been a couple of months, and so I just wanted to check in and see if there was anything else needed to move this along.

Also, Alan Malloy pointed out to me that my benchmarks above did not mention single-threaded performance. I actually wrote this into the tests above, but I neglected to report them at the time. Here are the results on the same machine as above (multithreaded versions are basically the same as the previous comment).

With a single thread performing the same work:
The lock-based multifns run at an average of 142ms.
The lockless multifns run at an average of 115ms.

So the lockless multimethods are still faster even in a single-threaded case, although the speedup is more modest compared to the speedups in the multithreaded cases above. This is not surprising, but it is good to know.

Screened. The approach is sound.

I can confirm similar performance measurements using David Santiago's benchmark, compared with Clojure 1.5 master as of commit f5f4faf.

Mean runtime (ms) of a multimethod when called repeatedly from N threads:

|            | N=1 | N=2 | N=4 |
|------------+-----+-----+-----|
| 1.5 master |  80 | 302 | 765 |
| lockless   |  63 |  88 | 125 |

My only concern is that the extra allocations of the State class will create more garbage, but this is probably not significant if we are already using persistent maps. It would be interesting to compare this implementation with one using thread-safe mutable data structures (e.g. ConcurrentHashMap) for the cache.

I think your assessment that it's not significant compared to the current implementation using persistent maps is correct. Regarding the extra garbage, note that the new State is only created when the hierarchy has changed or there's a cache miss (related, obviously)... situations where you're already screwed. Then it won't have to happen again for the same method (until another change to the multimethod). So for most code, it won't happen very often.

ConcurrentHashMap might be faster, it'd be interesting to see. My instinct is to keep it as close to being "made out of Clojure" as possible. In fact, it's hard to see why this couldn't be rewritten in Clojure itself some day, as I believe Chas Emerick has suggested. Also, I would point out that two of the three maps are used from the Clojure side in core.clj. I assume they would be happier if they were persistent maps.

Funny story: I was going to point out the parts of the code that were called from the clojure side just now, and alarmingly cannot find two of the functions. I think I must have misplaced them while rewriting the state into an immutable object. Going to attach a new patch with the fix and some tests for it in a few minutes.

Latest patch for this issue. Supersedes issue988-lockless-multifn+tests.diff as of 08/17/2012.

As promised, I reimplemented those two functions. I also added more multimethod tests to the test suite. The new tests should definitely prevent a similar mistake. While I was at it, I went through core.clj and found all the multimethod API functions I could and ensured that there were at least some basic functionality tests for all of them. The ones I found were: defmulti, defmethod, remove-all-methods, remove-method, prefer-method, methods, get-method, prefers (Some of those already had tests from the earlier version of the patch).

Really sorry about catching this right after you vetted the patch. 12771 test assertions were apparently not affected by prefers and methods ceasing to function, but now there are 12780 to help prevent a similar error. Since you just went through it, I'm leaving the older version of the patch up so you can easily see the difference to what I've added.

https://github.com/clojure/clojure/commit/83ebf814d5d6663c49c1b2d0d076b57638bff673 should fix these issues. The patch here was too much of a change to properly vet.

If you could though, I'd appreciate a patch with just the multimethod tests.

Patch clj-988-tests-only-patch-v1.txt dated Sep 15 2012 is a subset of David Santiago's
patch issue988-lockless-multifn+tests-120817.diff dated Aug 17 2012. It includes only the tests from that patch. Applies cleanly and passes tests with latest master after Rich's read/write lock change for multimethods was committed.

tests-only patch ok

Updated patch.

Applies properly against d4170e65d001c8c2976f1bd7159484056b9a9d6d. Things look good to me.