Observe the following program written in Java (complete runnable version follows, but the important part of the program is in the snippet a little bit further below):
import java.util.ArrayList;
/** A not easy to explain benchmark.
*/
class MultiVolatileJavaExperiment {
public static void main(String[] args) {
(new MultiVolatileJavaExperiment()).mainMethod(args);
}
int size = Integer.parseInt(System.getProperty("size"));
int par = Integer.parseInt(System.getProperty("par"));
public void mainMethod(String[] args) {
int times = 0;
if (args.length == 0) times = 1;
else times = Integer.parseInt(args[0]);
ArrayList < Long > measurements = new ArrayList < Long > ();
for (int i = 0; i < times; i++) {
long start = System.currentTimeMillis();
run();
long end = System.currentTimeMillis();
long time = (end - start);
System.out.println(i + ") Running time: " + time + " ms");
measurements.add(time);
}
System.out.println(">>>");
System.out.println(">>> All running times: " + measurements);
System.out.println(">>>");
}
public void run() {
int sz = size / par;
ArrayList < Thread > threads = new ArrayList < Thread > ();
for (int i = 0; i < par; i++) {
threads.add(new Reader(sz));
threads.get(i).start();
}
for (int i = 0; i < par; i++) {
try {
threads.get(i).join();
} catch (Exception e) {}
}
}
final class Foo {
int x = 0;
}
final class Reader extends Thread {
volatile Foo vfoo = new Foo();
Foo bar = null;
int sz;
public Reader(int _sz) {
sz = _sz;
}
public void run() {
int i = 0;
while (i < sz) {
vfoo.x = 1;
// with the following line commented
// the scalability is almost linear
bar = vfoo; // <- makes benchmark 2x slower for 2 processors - why?
i++;
}
}
}
}
Explanation : The program is actually very simple. It loads integers size
and par
from the system properties (passed to jvm with the -D
flag) - these are the input length and the number of threads to use later. It then parses the first command line argument which says how many time to repeat the program (we want to be sure that the JIT has done its work and have more reliable measurements).
The run
method is called in each repetition. This method simply starts par
threads, each of which will do a loop with size / par
iterations. The thread body is defined in the Reader
class. Each repetition of the loop reads a volatile member vfoo
and assigns 1
to its public field. After that, vfoo
is read once again and assigned to a non-volatile field bar
.
Notice how most of the time the program is executing the loop body, so the run
in the thread is the focus of this benchmark:
final class Reader extends Thread {
volatile Foo vfoo = new Foo();
Foo bar = null;
int sz;
public Reader(int _sz) {
sz = _sz;
}
public void run() {
int i = 0;
while (i < sz) {
vfoo.x = 1;
// with the following line commented
// the scalability is almost linear
bar = vfoo; // <- makes benchmark 2x slower for 2 processors - why?
i++;
}
}
}
Observations : Running java -Xmx512m -Xms512m -server -Dsize=500000000 -Dpar=1 MultiVolatileJavaExperiment 10
on an
Ubuntu Server 10.04.3 LTS
8 core Intel(R) Xeon(R) CPU X5355 @2.66GHz
~20GB ram
java version "1.6.0_26"
Java(TM) SE Runtime Environment (build 1.6.0_26-b03)
Java HotSpot(TM) 64-Bit Server VM (build 20.1-b02, mixed mode)
I get the following times:
>>> All running times: [821, 750, 1011, 750, 758, 755, 1219, 751, 751, 1012]
Now, setting -Dpar=2
, I get:
>>> All running times: [1618, 380, 1476, 1245, 1390, 1391, 1445, 1393, 1511, 1508]
Apparently, this doesn't scale for some reason - I would have expected the second output to be twice as fast (although it does seem to be in one of the early iterations - 380ms
).
Interestingly, commenting out the line bar = vfoo
(which isn't even supposed to be a volatile write), yields the following times for -Dpar
set to 1,2,4,8
.
>>> All running times: [762, 563, 563, 563, 563, 563, 570, 566, 563, 563]
>>> All running times: [387, 287, 285, 284, 283, 281, 282, 282, 281, 282]
>>> All running times: [204, 146, 143, 142, 141, 141, 141, 141, 141, 141]
>>> All running times: [120, 78, 74, 74, 81, 75, 73, 73, 72, 71]
It scales perfectly.
Analysis : First of all, there are no garbage collection cycles occuring here (I've added -verbose:gc
as well to check this).
I get similar results on my iMac.
Each thread is writing to its own field, and different Foo
object instances belonging to different threads don't appear to be ending up in the same cachelines - adding more members into Foo
to increase its size doesn't change the measurements. Each thread object instance has more than enough fields to fill up the L1 cache line. So this probably isn't a memory issue.
My next thought was that the JIT
might be doing something weird, because the early iterations usually do scale as expected in the uncommented version, so I checked this out by printing the assembly (see this post on how to do that ).
java -Xmx512m -Xms512m -server -XX:CompileCommand=print,*Reader.run MultiVolatileJavaExperiment -Dsize=500000000 -Dpar=1 10
and I get these 2 outputs for the 2 versions for the Jitted method run
in Reader
. The commented (properly scalable) version:
[Verified Entry Point]
0xf36c9fac: mov %eax,-0x3000(%esp)
0xf36c9fb3: push %ebp
0xf36c9fb4: sub $0x8,%esp
0xf36c9fba: mov 0x68(%ecx),%ebx
0xf36c9fbd: test %ebx,%ebx
0xf36c9fbf: jle 0xf36c9fec
0xf36c9fc1: xor %ebx,%ebx
0xf36c9fc3: nopw 0x0(%eax,%eax,1)
0xf36c9fcc: xchg %ax,%ax
0xf36c9fd0: mov 0x6c(%ecx),%ebp
0xf36c9fd3: test %ebp,%ebp
0xf36c9fd5: je 0xf36c9ff7
0xf36c9fd7: movl $0x1,0x8(%ebp)
---------------------------------------------
0xf36c9fde: mov 0x68(%ecx),%ebp
0xf36c9fe1: inc %ebx ; OopMap{ecx=Oop off=66}
;*goto
; - org.scalapool.bench.MultiVolatileJavaExperiment$Reader::run@21 (line 83)
---------------------------------------------
0xf36c9fe2: test %edi,0xf7725000 ; {poll}
0xf36c9fe8: cmp %ebp,%ebx
0xf36c9fea: jl 0xf36c9fd0
0xf36c9fec: add $0x8,%esp
0xf36c9fef: pop %ebp
0xf36c9ff0: test %eax,0xf7725000 ; {poll_return}
0xf36c9ff6: ret
0xf36c9ff7: mov $0xfffffff6,%ecx
0xf36c9ffc: xchg %ax,%ax
0xf36c9fff: call 0xf36a56a0 ; OopMap{off=100}
;*putfield x
; - org.scalapool.bench.MultiVolatileJavaExperiment$Reader::run@15 (line 79)
; {runtime_call}
0xf36ca004: call 0xf6f877a0 ; {runtime_call}
The uncommented bar = vfoo
(non-scalable, slower) version:
[Verified Entry Point]
0xf3771aac: mov %eax,-0x3000(%esp)
0xf3771ab3: push %ebp
0xf3771ab4: sub $0x8,%esp
0xf3771aba: mov 0x68(%ecx),%ebx
0xf3771abd: test %ebx,%ebx
0xf3771abf: jle 0xf3771afe
0xf3771ac1: xor %ebx,%ebx
0xf3771ac3: nopw 0x0(%eax,%eax,1)
0xf3771acc: xchg %ax,%ax
0xf3771ad0: mov 0x6c(%ecx),%ebp
0xf3771ad3: test %ebp,%ebp
0xf3771ad5: je 0xf3771b09
0xf3771ad7: movl $0x1,0x8(%ebp)
-------------------------------------------------
0xf3771ade: mov 0x6c(%ecx),%ebp
0xf3771ae1: mov %ebp,0x70(%ecx)
0xf3771ae4: mov 0x68(%ecx),%edi
0xf3771ae7: inc %ebx
0xf3771ae8: mov %ecx,%eax
0xf3771aea: shr $0x9,%eax
0xf3771aed: movb $0x0,-0x3113c300(%eax) ; OopMap{ecx=Oop off=84}
;*goto
; - org.scalapool.bench.MultiVolatileJavaExperiment$Reader::run@29 (line 83)
-----------------------------------------------
0xf3771af4: test %edi,0xf77ce000 ; {poll}
0xf3771afa: cmp %edi,%ebx
0xf3771afc: jl 0xf3771ad0
0xf3771afe: add $0x8,%esp
0xf3771b01: pop %ebp
0xf3771b02: test %eax,0xf77ce000 ; {poll_return}
0xf3771b08: ret
0xf3771b09: mov $0xfffffff6,%ecx
0xf3771b0e: nop
0xf3771b0f: call 0xf374e6a0 ; OopMap{off=116}
;*putfield x
; - org.scalapool.bench.MultiVolatileJavaExperiment$Reader::run@15 (line 79)
; {runtime_call}
0xf3771b14: call 0xf70307a0 ; {runtime_call}
The differences in the two versions are within ---------
. I expected to find synchronization instructions in the assembly which might account for the performance issue - while few extra shift
, mov
and inc
instructions might affect absolute performance numbers, I don't see how they could affect scalability.
So, I suspect that this is some sort of a memory issue related to storing to a field in the class. On the other hand, I'm also inclined to believe that the JIT does something funny, because in one iteration the measured time is twice as fast, as it should be.
Can anyone explain what is going on here? Please be precise and include references that support your claims.
Thank you!
EDIT:
Here's the bytecode for the fast (scalable) version:
public void run();
LineNumberTable:
line 77: 0
line 78: 2
line 79: 10
line 83: 18
line 85: 24
Code:
Stack=2, Locals=2, Args_size=1
0: iconst_0
1: istore_1
2: iload_1
3: aload_0
4: getfield #7; //Field sz:I
7: if_icmpge 24
10: aload_0
11: getfield #5; //Field vfoo:Lorg/scalapool/bench/MultiVolatileJavaExperiment$Foo;
14: iconst_1
15: putfield #8; //Field org/scalapool/bench/MultiVolatileJavaExperiment$Foo.x:I
18: iinc 1, 1
21: goto 2
24: return
LineNumberTable:
line 77: 0
line 78: 2
line 79: 10
line 83: 18
line 85: 24
StackMapTable: number_of_entries = 2
frame_type = 252 /* append */
offset_delta = 2
locals = [ int ]
frame_type = 21 /* same */
The slow (non-scalable) version with bar = vfoo
:
public void run();
LineNumberTable:
line 77: 0
line 78: 2
line 79: 10
line 82: 18
line 83: 26
line 85: 32
Code:
Stack=2, Locals=2, Args_size=1
0: iconst_0
1: istore_1
2: iload_1
3: aload_0
4: getfield #7; //Field sz:I
7: if_icmpge 32
10: aload_0
11: getfield #5; //Field vfoo:Lorg/scalapool/bench/MultiVolatileJavaExperiment$Foo;
14: iconst_1
15: putfield #8; //Field org/scalapool/bench/MultiVolatileJavaExperiment$Foo.x:I
18: aload_0
19: aload_0
20: getfield #5; //Field vfoo:Lorg/scalapool/bench/MultiVolatileJavaExperiment$Foo;
23: putfield #6; //Field bar:Lorg/scalapool/bench/MultiVolatileJavaExperiment$Foo;
26: iinc 1, 1
29: goto 2
32: return
LineNumberTable:
line 77: 0
line 78: 2
line 79: 10
line 82: 18
line 83: 26
line 85: 32
StackMapTable: number_of_entries = 2
frame_type = 252 /* append */
offset_delta = 2
locals = [ int ]
frame_type = 29 /* same */
The more I am experimenting with this, it seems to me that this has nothing to do with volatiles at all - it has something to do with writing to object fields. My hunch is that this is somehow a memory contention issue - something with caches and false sharing, although there is no explicit synchronization at all.
EDIT 2:
Interestingly, changing the program like this:
final class Holder {
public Foo bar = null;
}
final class Reader extends Thread {
volatile Foo vfoo = new Foo();
Holder holder = null;
int sz;
public Reader(int _sz) {
sz = _sz;
}
public void run() {
int i = 0;
holder = new Holder();
while (i < sz) {
vfoo.x = 1;
holder.bar = vfoo;
i++;
}
}
}
resolves the scaling issue. Apparently, the Holder
object above gets created after the thread is started, and is probably allocated in a different segment of memory, which is then being modified concurrently, as opposed to modifying the field bar
in the thread object, which is somehow "close" in memory between different thread instances.
This is what I think is happening (keep in mind I'm not familiar with HotSpot):
ReplyDelete0xf36c9fd0: mov 0x6c(%ecx),%ebp ; vfoo
0xf36c9fd3: test %ebp,%ebp ; vfoo is null?
0xf36c9fd5: je 0xf36c9ff7 ; throw NullPointerException (I guess)
0xf36c9fd7: movl $0x1,0x8(%ebp) ; vfoo.x = 1
0xf36c9fde: mov 0x68(%ecx),%ebp ; sz
0xf36c9fe1: inc %ebx ; i++
0xf36c9fe2: test %edi,0xf7725000 ; safepoint on end of loop
0xf36c9fe8: cmp %ebp,%ebx ; i < sz?
0xf36c9fea: jl 0xf36c9fd0
0xf3771ad0: mov 0x6c(%ecx),%ebp ; vfoo
0xf3771ad3: test %ebp,%ebp ; vfoo is null?
0xf3771ad5: je 0xf3771b09 ; throw NullPointerException (I guess)
0xf3771ad7: movl $0x1,0x8(%ebp) ; vfoo.x = 1
0xf3771ade: mov 0x6c(%ecx),%ebp ; \
0xf3771ae1: mov %ebp,0x70(%ecx) ; / bar = vfoo
0xf3771ae4: mov 0x68(%ecx),%edi ; sz
0xf3771ae7: inc %ebx ; i++
0xf3771ae8: mov %ecx,%eax ;
0xf3771aea: shr $0x9,%eax ; ??? \ Probably replaced later
0xf3771aed: movb $0x0,-0x3113c300(%eax) ; ??? / by some barrier code?
0xf3771af4: test %edi,0xf77ce000 ; safepoint
0xf3771afa: cmp %edi,%ebx ; i < sz ?
0xf3771afc: jl 0xf3771ad0 ;
The reason I think the above code stands in for a barrier is that when taking the NullPointerException, the scalable version has a XCHG, which acts as a barrier, while the non-scalable version has a NOP there.
The rationale would be that there needs to be a happens-before ordering between the initial load of vfoo and joining the thread. In the volatile case, the barrier would be inside the loop, so it wouldn't need to be elsewhere. What I don't understand is why XCHG isn't used inside the loop. Maybe runtime detection of MFENCE support?
You are not actually writing to a volatile field so the volatile field can be cached in each thread.
ReplyDeleteUsing volatile prevents some compiler optimisations and in a micro-benchmark, you can see a large relative difference.
In the example above, the commented out version is longer because it has loop unrolled to place two iterations in one actual loop. This can almost double performance.
When using volatile you can see there is no loop unrolling.
BTW: You can remove a lot of the code in your example to make it easier to read. ;)
Edit: This answer did not stand up to testing.
ReplyDeleteI have no way to test this right now (no multicore CPU in this machine), but here is a theory: The Foo instances might not be in the same cache lines, but perhaps the Reader instances are.
This means the slowdown could be explained by the write to bar, rather than the read of foo, because writing to bar would invalidate that cache line for the other core and cause lots of copying between caches. Commenting out the write to bar (which is the only write to a field of Reader in the loop) stops the slowdown, which is consistent with this explanation.
Edit: According to this article, the memory layout of objects is such that the bar reference would be the last field in the layout of the Reader object. This means it is probable to land in the same cache line as the next object on the Heap. Since I am not sure about the order in which new objects are allocated on the Heap, I suggested in the comment below to pad both "hot" object types with references, which would be effective in separating the objects (At least, I hope it will, but it depends on how fields of the same type are sorted in memory).