What I did that helped is that I grabbed the ConcurrentQueue code from corefx and added it locally to the project, and run the profiler.
We can see that in the code that is on github, ~30 % of the time is spent on TryDequeue and 30% is on Count
most of this time is just looping through segments inside the concurrent queue
so, it's not actually locks, just regular code running. it's hard to optimize it more for general use cases.
(when I say locks in concurrentqueue, I mean spinwait, not the lock keyword)
if you were to pass all the overhead to adding and zero to retrieving, the most optimal solution would be to use a simple List<T> with a predefined size using locks/spinwait on Add, and for dequeue you have a single method that locks/spinwait, and exchange the list with a new one. you can then loop the list with tasks in the pool directly as an array. the issue is that it's hard to know what size this array should be
so, becomes a trade off to a linkedlist doing the same.
after that, the path would be a linked list of arrays, and that's exactly what ConcurrentQueue is...
the point I said about the test is that unless you have some time to fill a list with some items without dequeueing, the gains you have for swapping the list are lost. in the test, it keeps producing and swapping all the time, so the cost becomes of swapping becomes higher. if you just give it some time to fill a list and swap with a minimum amount of items, swapping becomes cheaper and these other solutions apply
another thing.. if using concurrentqueue, replacing Count for IsEmpty is much cheaper, as IsEmpty only checks the local segment, while Count needs to loop through all of them
another solution would be to add a true
DequeAll in a modified version of ConcurrentQueue. that would be feasable too, right? we could just make the concurrentqueue head point to where the last segment is, and then enumerate over all of the segments we just removed... havent checked the code for it, but that should work, right?
as I understand, what it does inside is that it creates a 32-item array for each segment, and it grows by linking them. Each segment has its lock, so you only lock if you are modifying that 32-item boundary
From the profiler, even with the 4th iteration (10 million actions I guess?), the time it wasted on spinwait was about 4% only. 30% was just moving through segments
so, in theory reducing the # of segments (that is - increasing the segment size) could help...
but at some point I wonder if you're not optimizing only for the benchmark
in the case of DequeueAll, if it needs to loop and lock through all segments, there might not be a lot of gain. But it could get the entire segment as an array and work those before going to the next
1-2 weeks IIRC
and for surrogates, surrogate support is mandatory as
ActorRef needs special handling to be resolved into a real actorref.. and it can be embedded deeply inside another message
I do want to get rid of Json.NET, but most other serializers lack some features
e.g. polymorphic serialization
and what about 2nd? Don't you think we should add a globally configurable HOCON option for this? I think it's very risky that now we don't have any kind of timeout configured by default. It's not safe, if user forgot at any point to set receive timeout, his/her service could possibly hang forever
if you mean a way to ensure that no receive method runs forever.. I have looked into this, but there is no fast way to deal with it. there would need to be some supervisor thread that polls that no thread have been executing a receiveblock for too long
During the very early stages of akka.net, I put together this: https://gist.github.com/rogeralsing/8472797 it does what you want, but expensive, and the implementation would beed to be cleaned
but eventually I just moved it to separate spec class
I've seen a similar problem before, back then, it was that some static resources, like NoRouter or DefaultDeploy or whatver, was overwritten by the serializer .. e.g. some primitives took a static resource as an input argument, then the serializer overwrote the properties of the primitive, which was backed by the static resource, and tada , the static resource was corrupted
could that be a similar problem here?
e.g.
public Props() : this(Some.Static.Resource) {} then when the deserializer tried to deserialize an object with such ctor, the injected resource would be corrupted