digitalmars.D - GC experiments. Writing my own GC.
- Adam Sakareassen via Digitalmars-d (105/105) May 12 2014 Hi all,
- FrankLike (4/6) May 12 2014 Congratulations!
- Rainer Schuetze (18/50) May 13 2014 [...]
- Adam Sakareassen via Digitalmars-d (6/64) May 13 2014 Destructors will almost certainly be run on the background thread.
- Andrei Alexandrescu (4/9) May 13 2014 [snip]
- Adam Sakareassen via Digitalmars-d (10/20) May 13 2014 Thank-you.
- =?UTF-8?B?Ik5vcmRsw7Z3Ig==?= (3/5) May 30 2014 This sounds almost to good to be true!
- Wanderer (2/8) May 30 2014 Probably this is the reason why results are so nice :-)
Hi all,
As a learning exercise I've just been doing some experimenting with
rewriting the garbage collection code, and thought I might share some of
the initial results. I only program as a hobby these days, and I'm
certainly no expert, but I thought some people might find it interesting.
My interest started because I wrote a LR1 file parser in D. I then
multi threaded the application so multiple files could be parsed
simultaneously. (Disk IO was all on one thread). To my surprise, the
through-put dropped significantly. I could process the files a lot
faster using only one thread. It turns out the delays were due to my
liberal use of the “new” statement. Rather than block allocate the
memory I thought I would just hack the GC. So I cleared out gc.d in the
runtime and started again. The basic plan was to make it more
multi-thread friendly. Already I have learnt quite a lot. There are a
number of things I would do differently if I tried another re-write.
However so far it has been a good learning experience.
Currently, allocations are all working, and the mark phase is running.
It still will not sweep and free the memory.
All memory allocations are entirely lock free (using CAS instructions).
So a pre-empted thread will never block another. For allocations of
less than 128 bytes, each thread is allocated memory from it's own
memory pool to avoid false sharing on the CPU's cache. The collector
component runs on a background thread using a mark and sweep algorithm
which is basically the same as the existing algorithm. Currently the
thread will wake up every 100ms and decide if a collection should be
performed. An emergency collection will run in the foreground if a
memory allocation fails during that period.
The mark phase needs to stop the world. The sweeping portion of the
collection will run in the background. This is similar to the current
implementation as the world is restarted after the mark phase, however
the thread doing the collection will not allocate the requested memory
to the calling thread until after the sweep has completed. This means
that single threaded applications always wait for the full garbage
collection cycle.
So far allocation speed seems to have improved. I can't test collection
speed as it's not complete. As a test I wrote a simple function that
allocates a linked list of 2 million items. This function is then
spawned by 20 threads. This test script is shown below. Timing for
allocation (with GC disabled) is as follows. (Using DMD 2.065)
Existing GC code: 15700ms (average)
My GC code: 500ms (Average)
When performing the same amount of allocations on a single thread, the
new code is still slightly faster than the old.
What this demonstrates is that the locking mechanisms in the current GC
code is a huge overhead for multi threaded applications that perform a
lot of memory allocations. (ie. Use the “new” operator or dynamic
arrays.)
It would be nice to see the default GC and memory allocator improved.
There is certainly room for improvement on the allocator end which may
mask some of the performance issues associated with garbage collection.
In the future I think D needs to look at making collection precise. It
would not be too hard to adjust the mark and sweep GC to be nearly
precise. The language needs to support precise GC before things like
moving garbage collection become feasible.
Anyway, I just thought I'd share the results of my experimenting. I
would be happy to make the code available in a few weeks time. Perhaps
someone might find is useful. I need to get it finished and tested
first. :-)
Cheers!
Adam
------
//Test script that generated these results:
import std.stdio;
import std.datetime;
import std.concurrency;
import core.memory;
class LinkedList{
long value =0;
LinkedList next;
}
shared int threadCount = 0;
void main(){
core.memory.GC.disable();
auto start = Clock.currSystemTick();
foreach(i; 0 .. 20){
auto tid = spawn(&doSomething, thisTid);
threadCount++;
}
while(threadCount >0){};
auto ln = Clock.currSystemTick() - start;
writeln(ln.msecs, "ms");
}
void doSomething(Tid tid){
auto top = new LinkedList;
auto recent = top;
//Create the linked list
foreach(i; 1 .. 2_000_000){
auto newList = new LinkedList;
newList.value = i;
recent.next = newList;
recent = newList;
}
//Sum the values. (Just spends some time walking the memory).
recent = top;
long total=0;
while(recent !is null){
total += recent.value;
recent = recent.next;
}
writeln("Total : ", total );
threadCount--;
}
May 12 2014
"FrankLike" <1150015857 qq.com> Existing GC code: 15700ms (average) My GC code: 500ms (Average)Congratulations! Can you share you good work for us? or exe? dll? Thank you. Frank
May 12 2014
This sounds pretty cool. On 13.05.2014 06:42, Adam Sakareassen via Digitalmars-d wrote:Hi all,[...]All memory allocations are entirely lock free (using CAS instructions). So a pre-empted thread will never block another. For allocations of less than 128 bytes, each thread is allocated memory from it's own memory pool to avoid false sharing on the CPU's cache.There was a GSoC project a few years ago which was planning to do something similar, but then switched to go for implementing precise scanning. I always wanted to add lock-free memory allocations, too. [...]The mark phase needs to stop the world. The sweeping portion of the collection will run in the background. This is similar to the current implementation as the world is restarted after the mark phase, however the thread doing the collection will not allocate the requested memory to the calling thread until after the sweep has completed. This means that single threaded applications always wait for the full garbage collection cycle.Does this mean that destructors/finalizers are called from a different thread? It was never guaranteed to be run on the same thread as the allocation of the object so far, but effectively that happens for single-threaded applications. It could help to continue doing that, some resources have to be released from the same thread they have been acquired (e.g. UI handles). For multi-threaded applications, I'm fine with "guaranteeing" that the destructor is never run on the same thread, though some standard mechanism should exist to forward to another thread.So far allocation speed seems to have improved. I can't test collection speed as it's not complete. As a test I wrote a simple function that allocates a linked list of 2 million items. This function is then spawned by 20 threads. This test script is shown below. Timing for allocation (with GC disabled) is as follows. (Using DMD 2.065) Existing GC code: 15700ms (average) My GC code: 500ms (Average)Very promising numbers :-)When performing the same amount of allocations on a single thread, the new code is still slightly faster than the old. What this demonstrates is that the locking mechanisms in the current GC code is a huge overhead for multi threaded applications that perform a lot of memory allocations. (ie. Use the “new” operator or dynamic arrays.) It would be nice to see the default GC and memory allocator improved. There is certainly room for improvement on the allocator end which may mask some of the performance issues associated with garbage collection. In the future I think D needs to look at making collection precise. It would not be too hard to adjust the mark and sweep GC to be nearly precise. The language needs to support precise GC before things like moving garbage collection become feasible.You might check how compatible your implementation is with this: https://github.com/rainers/druntime/tree/gcx_precise2
May 13 2014
Destructors will almost certainly be run on the background thread. Although those run at shutdown will likely be on the main thread. I hope the overall performance results are good. I really need to see the timing of the collections to know. But in any case it does indicate that some work on removing lock contention is worthwhile. On 13/05/2014 5:04 PM, Rainer Schuetze via Digitalmars-d wrote:This sounds pretty cool. On 13.05.2014 06:42, Adam Sakareassen via Digitalmars-d wrote:Hi all,[...]All memory allocations are entirely lock free (using CAS instructions). So a pre-empted thread will never block another. For allocations of less than 128 bytes, each thread is allocated memory from it's own memory pool to avoid false sharing on the CPU's cache.There was a GSoC project a few years ago which was planning to do something similar, but then switched to go for implementing precise scanning. I always wanted to add lock-free memory allocations, too. [...]The mark phase needs to stop the world. The sweeping portion of the collection will run in the background. This is similar to the current implementation as the world is restarted after the mark phase, however the thread doing the collection will not allocate the requested memory to the calling thread until after the sweep has completed. This means that single threaded applications always wait for the full garbage collection cycle.Does this mean that destructors/finalizers are called from a different thread? It was never guaranteed to be run on the same thread as the allocation of the object so far, but effectively that happens for single-threaded applications. It could help to continue doing that, some resources have to be released from the same thread they have been acquired (e.g. UI handles). For multi-threaded applications, I'm fine with "guaranteeing" that the destructor is never run on the same thread, though some standard mechanism should exist to forward to another thread.So far allocation speed seems to have improved. I can't test collection speed as it's not complete. As a test I wrote a simple function that allocates a linked list of 2 million items. This function is then spawned by 20 threads. This test script is shown below. Timing for allocation (with GC disabled) is as follows. (Using DMD 2.065) Existing GC code: 15700ms (average) My GC code: 500ms (Average)Very promising numbers :-)When performing the same amount of allocations on a single thread, the new code is still slightly faster than the old. What this demonstrates is that the locking mechanisms in the current GC code is a huge overhead for multi threaded applications that perform a lot of memory allocations. (ie. Use the “new” operator or dynamic arrays.) It would be nice to see the default GC and memory allocator improved. There is certainly room for improvement on the allocator end which may mask some of the performance issues associated with garbage collection. In the future I think D needs to look at making collection precise. It would not be too hard to adjust the mark and sweep GC to be nearly precise. The language needs to support precise GC before things like moving garbage collection become feasible.You might check how compatible your implementation is with this: https://github.com/rainers/druntime/tree/gcx_precise2
May 13 2014
On 5/12/14, 9:42 PM, Adam Sakareassen via Digitalmars-d wrote:Hi all, As a learning exercise I've just been doing some experimenting with rewriting the garbage collection code, and thought I might share some of the initial results. I only program as a hobby these days, and I'm certainly no expert, but I thought some people might find it interesting.[snip] This all is _very_ encouraging, please do share. Better yet, please contribute patches (one bite-size at a time) to dmd!! -- Andrei
May 13 2014
Thank-you. Once I am happy with result I will certainly share. The internal structures of the GC have been redesigned so I can't simply patch the existing file. Once it is fully operational I will make it available for testing. The good performance numbers are for a very specific test case, so it would be nice to have some more generalised testing once we get there. I would welcome the work being included in DMD if others thought it suited the generalised case. On 13/05/2014 5:52 PM, Andrei Alexandrescu via Digitalmars-d wrote:On 5/12/14, 9:42 PM, Adam Sakareassen via Digitalmars-d wrote:Hi all, As a learning exercise I've just been doing some experimenting with rewriting the garbage collection code, and thought I might share some of the initial results. I only program as a hobby these days, and I'm certainly no expert, but I thought some people might find it interesting.[snip] This all is _very_ encouraging, please do share. Better yet, please contribute patches (one bite-size at a time) to dmd!! -- Andrei
May 13 2014
Existing GC code: 15700ms (average) My GC code: 500ms (Average)This sounds almost to good to be true! What remains to be fixed before a pull request can be made? /Per
May 30 2014
On Friday, 30 May 2014 at 10:45:14 UTC, Nordlöw wrote:Existing GC code: 15700ms (average) My GC code: 500ms (Average)This sounds almost to good to be true!Probably this is the reason why results are so nice :-)For allocations of less than 128 bytes, each thread is allocated memory from it's own memory pool to avoid false sharing on the CPU's cache.
May 30 2014