VMs should share GC threads
https://bugs.webkit.org/show_bug.cgi?id=149433
rdar://problem/12859344
Reviewed by Geoffrey Garen.
Source/JavaScriptCore:
This changes the GC to use a new WTF abstraction for parallelism called ParallelHelperPool.
This allows us to remove GCThread and all of the GCPhase machinery. This kills a lot of
code and also gives our GC magical thread sharing powers. If two GCs in two different VMs
fire at the same time, then they will both get a random subset of the available shared GC
threads. If one GC happens before the other, then it will probably get all of the available
threads. If a GC happens while another VM already started GCing, then it will probably not
get any helper threads. This is probably fine, since in multi-VM scenarios we have no
reason to optimize for anything other than total throughput.
The GC has one static helper pool. This pool is available via JSC::heapHelperPool(). It
would be OK for other parts of JSC to use it in the future for parallel tasks. Each Heap
instance has a helper client attached to the pool.
The marking phase tells the ParallelHelperClient to asynchronously run a function that
joins parallel marking and finishes once marking reaches termination. It uses the
client.setFunction() idiom where the threads share work with each other using a specialized
worklist. The ParallelHelperPool is not involved in deciding when threads should terminate.
The copying phase tells the ParallelHelperClient to run a copying function in parallel. It
uses the client.runFunctionInParallel() idiom. The copying function gets work from the
m_blocksToCopy worklist inside Heap.
To test that multiple VMs work properly, this adds a multi-VM test to testapi.mm. This test
creates five concurrent VMs and has each of them allocate about 30MB of memory before doing
a full GC. I've confirmed that this tests uses only 6 total GC threads on my 8-core
computer (this is correct since we are currently configured for 7-way parallelism).
This shouldn't affect performance on benchmarks, but it will sure help apps with a lot of
VM instances.
- CMakeLists.txt:
- JavaScriptCore.vcxproj/JavaScriptCore.vcxproj:
- JavaScriptCore.vcxproj/JavaScriptCore.vcxproj.filters:
- JavaScriptCore.xcodeproj/project.pbxproj:
- assembler/AbstractMacroAssembler.h:
- heap/GCThread.cpp: Removed.
- heap/GCThread.h: Removed.
- heap/Heap.cpp:
(JSC::Heap::Heap):
(JSC::Heap::~Heap):
(JSC::Heap::isPagedOut):
(JSC::Heap::markRoots):
(JSC::Heap::copyBackingStores):
(JSC::Heap::resetVisitors):
(JSC::Heap::threadVisitCount):
(JSC::Heap::threadBytesVisited):
(JSC::Heap::threadBytesCopied):
(JSC::Heap::startNextPhase): Deleted.
(JSC::Heap::endCurrentPhase): Deleted.
- heap/Heap.h:
- heap/HeapHelperPool.cpp: Added.
(JSC::heapHelperPool):
- heap/HeapHelperPool.h: Added.
- heap/MarkStack.cpp:
(JSC::MarkStackArray::stealSomeCellsFrom):
(JSC::SlotVisitor::didStartMarking):
(JSC::SlotVisitor::reset):
(JSC::SlotVisitor::drainFromShared):
- jit/BinarySwitch.h:
- runtime/CodeCache.h:
- runtime/VM.h:
- runtime/WeakRandom.h: Removed.
- API/tests/testapi.mm:
Source/WTF:
This adds two major things to WTF: WeakRandom and ParallelHelperPool. WeakRandom was
already in JSC; we're just hoisting it into WTF. It's just a weak random number generator
that's suitable for places where you need just a tiny bit of randomness.
ParallelHelperPool is a new API that simplifies data-parallel algorithms like the JSC GC.
In a data-parallel algorithm, we want to run one task on as many cores as possible and let
the task worry about which subset of the input data to work on. In some cases, the
algorithm will not need to do any load balancing - and if load balancing is required, it's
up to the user. This is appropriate in contexts where the load balancing needs to be
custom-tuned for performance, like the GC's marking phase.
This new API has three concepts: task, client, and pool. A task is a reference counted
object with a run() method, which may be run in parallel. It is usually used to wrap a
functor. A pool is a pool of threads that can run things. A client is a placeholder for a
task. A client can have zero or one tasks. A client must be registered with a pool. When a
client has a task, the pool's threads may choose to run it. If a thread starts running a
task, it will run it to completion. When the task returns on any thread, the client takes
it to mean that the task should be removed. That means that any currently running instances
of the task will finish but no new threads will attempt to run the task. You can easily ask
a client to wait until a task finishes. You can also easily ask a client to run a task on
the current thread in addition to possibly some helper threads from the pool.
For some data-parallel algorithms, programming with ParallelHelperPool is as easy as:
client.runFunctionInParallel(
[=] () {
do things;
});
Note that you cannot tell ahead of time how many threads will join to help the task.
Threads may become available after the task has already started running. Those threads may
join after the other threads have already started. It's not advisable to make algorithmic
decisions based on client.numberOfActiveThreads(), since that number may change. Usually
the best way to use ParallelHelperPool is with an algorithm that has its own custom
worklist. An example of a very simple custom worklist is the one in the JSC GC's copying
phase - it's just a Vector and an index that indicates the next set of elements to process.
This new API was initially designed to simplify how GCThread works, by replacing Phase with
a callback that contains the phase's workload. I then realized that with a few tweaks, I
could make this somewhat general enough that it might become interesting outside GC. I also
realized that I could use this to enable thread sharing. So, although the API is kinda
quirky, it's grounded in the reality of how the JSC GC does parallelism.
- WTF.vcxproj/WTF.vcxproj:
- WTF.vcxproj/WTF.vcxproj.filters:
- WTF.xcodeproj/project.pbxproj:
- wtf/CMakeLists.txt:
- wtf/ParallelHelperPool.cpp: Added.
(WTF::ParallelHelperClient::ParallelHelperClient):
(WTF::ParallelHelperClient::~ParallelHelperClient):
(WTF::ParallelHelperClient::setTask):
(WTF::ParallelHelperClient::finish):
(WTF::ParallelHelperClient::doSomeHelping):
(WTF::ParallelHelperClient::runTaskInParallel):
(WTF::ParallelHelperClient::claimTask):
(WTF::ParallelHelperClient::runTask):
(WTF::ParallelHelperPool::ParallelHelperPool):
(WTF::ParallelHelperPool::~ParallelHelperPool):
(WTF::ParallelHelperPool::addThreads):
(WTF::ParallelHelperPool::ensureThreads):
(WTF::ParallelHelperPool::doSomeHelping):
(WTF::ParallelHelperPool::didMakeWorkAvailable):
(WTF::ParallelHelperPool::helperThreadBody):
(WTF::ParallelHelperPool::hasClientWithTask):
(WTF::ParallelHelperPool::getClientWithTask):
(WTF::ParallelHelperPool::waitForClientWithTask):
- wtf/ParallelHelperPool.h: Added.
(WTF::ParallelHelperClient::setFunction):
(WTF::ParallelHelperClient::runFunctionInParallel):
(WTF::ParallelHelperClient::pool):
(WTF::ParallelHelperClient::numberOfActiveThreads):
(WTF::ParallelHelperPool::numberOfThreads):
(WTF::SharedTask::SharedTask):
(WTF::SharedTask::~SharedTask):
(WTF::SharedTaskFunctor::SharedTaskFunctor):
(WTF::createSharedTask):
- wtf/WeakRandom.h: Copied from Source/JavaScriptCore/runtime/WeakRandom.h.
(WTF::WeakRandom::WeakRandom):
(WTF::WeakRandom::initializeSeed):
(WTF::WeakRandom::seedUnsafe):
(WTF::WeakRandom::getUint32):
(WTF::WeakRandom::advance):
(JSC::WeakRandom::WeakRandom): Deleted.
(JSC::WeakRandom::seedUnsafe): Deleted.
(JSC::WeakRandom::getUint32): Deleted.
(JSC::WeakRandom::advance): Deleted.
(JSC::WeakRandom::initializeSeed): Deleted.
