I was trying to find some resources for best performance and scaling with message passing. I heard that message passing by value instead of reference can be better scalability as it works well with NUMA style setups and reduced contention for a given memory address.
I would assume value based message passing only works with "smaller" messages. What would "smaller" be defined as? At what point would references be better? Would one do stream processing this way?
I'm looking for some helpful tips or resources for these kinds of questions.
Thanks :-)
P.S. I work in C#, but I don't think that matters so much for these kind of design questions.
Some factors to add to the excellent advice of Jeremy:
1) Passing by value only works efficiently for small messages. If the data has a [cache-line-size] unused area at the start to avoid false sharing, you are already approaching the size where passing by reference is more efficient.
2) Wider queues mean more space taken up by the queues, impacting memory use.
3) Copying data into/outof wide queue structures takes time. Apart from the actual CPU use while moving data, the queue remains locked during the copying. This increases contention on the queue and leading to an overall performance hit that is queue width dependent. If there is any deadlock-potential in your code, keeping locks for extended periods will not help matters.
4) Passing by value tends to lead to code that is specific to the data size, ie. is fixed at compile-time. Apart from a nasty infestation of templates, this makes it very difficult to tune buffer-sizes etc. at run-time.
5) If the messages are passed by reference and malloced/freed/newed/disposed/GC'd, this can lead to excessive contention on the memory-manager and frequent, wasteful GC. I usually use fixed pools of messages, allocated at startup, specifically to avoid this.
6) Handling byte-streams can be awkward when passing by reference. If a byte-stream is characterized by frequent delivery of single bytes, pass-by-reference is only sensible if the bytes are chunked-up. This can lead to the need for timeouts to ensure that partially-filled messages are dispatched to the next thread in a timely manner. This introduces complication and latency.
7) Pass-by-reference designs are inherently more likely to leak. This can lead to extended test times and overdosing on valgrind - a particularly painful addiction, (another reason I use fixed-size message object pools).
8) Complex messages, eg. those that contain references to other objects, can cause horrendous problems with ownership and lifetime-management if passed by value. Example - a server socket object has a reference to a buffer-list object that contains an array of buffer-instances of varying size, (real example from IOCP server). Try passing that by value..
9) Many OS calls cannot handle anything but a pointer. You cannot PostMessage, (that's a Windows API, for all you happy-feet), even a 256-byte structure by value with one call, (you have just the 2 wParam,lParam integers). Calls that set up asychronous callbacks often allow 'context data' to be sent to the callback - almost always just one pointer. Any app that is going to use such OS functionality is almost forced to resort to pass by reference.
Jeremy Friesner's comment seems to be the best as this is a new area, although Martin James's points are also good. I know Microsoft is looking into message passing for their future kernels as we gain more cores.
There seems to be a framework that deals with message passing and it claims to have much better performance than current .Net producer/consumer generics. I'm not sure how it will compare to .Net's Dataflow in 4.5
https://github.com/odeheurles/Disruptor-net
Related
I'm successfully using a mpsc::channel() to send messages from a producer thread to a consumer.
The consumer is only ever interested in the latest message. (It uses the message from the previous check if there is no new message.)
In consequence, I'm running the consumer's try_recv() in a loop until it fails to get a new message, and then using the last received message, or the old one if no new messages were found.
Memory is being wasted storing old messages which the consumer will throw away.
How would I build a one-element variant of mpsc::channel()?
(I've considered using sync::Mutex<Option<MyMessage>> but it is critical that the consuming thread blocks for as little time as possible. Also, I want ownership to pass from the producer to the consumer.)
You can do it with an AtomicPtr, whose compare_exchange method should compile to a simple cmpxchg instruction, allowing you to store either std::ptr::null or an actual message.
There's quite a few possibilities, with various trade-offs.
I'd recommend the arc-swap crate (see below) for a safe and fast interface, and the DIY Double Buffering approach if performance is that critical.
std::mpsc
There's a second option for std::mpsc: the sync_channel function creates a bounded channel, where the sender blocks when the channel is full, until the receiver picks off a message.
I do not think that it is ideal for your usecase.
Tokio Watch channel
The Tokio ecosystem has the watch channel designed for the purpose of propagating configuration changes.
Unfortunately it is designed for multiple consumers, so the consumers borrow the messages: there is no transfer of ownership.
Arc Swap
I believe the arc-swap crate may be closer to what you need. As the name implies, it provides the moral equivalent of an Atomic<Arc<T>>.
You can use the ArcSwapOption<T> to have the equivalent of an Atomic<Option<Arc<T>>>, and the consumer can simply perform a let new = atomic.swap(None); then check if new is None (nothing new) or Some(Arc<T>) in which case it received an updated configuration.
Do be mindful of the cost of the dropping the previous Arc<T> when swapping a new one in: free is typically more expensive than malloc.
Back to std
You could use an AtomicPtr<T>. It'll require you to use unsafe, and would be a smidgen faster than ArcSwap by virtue of avoiding the reference counting.
It would suffer from the same drop issue, though.
DIY Double Buffering
You could also simply Do It Yourself. A simple double-buffering storage would work.
By storing a plain Option<T>, you avoid the additional extra allocation (and thus extra de-allocation), at the cost of making the check itself slower -- as you may now need to check both buffers. It may be possible to check a single buffer, not clear.
I want to see the intrinsic difference between a thread and a long-running go block in Clojure. In particular, I want to figure out which one I should use in my context.
I understand if one creates a go-block, then it is managed to run in a so-called thread-pool, the default size is 8. But thread will create a new thread.
In my case, there is an input stream that takes values from somewhere and the value is taken as an input. Some calculations are performed and the result is inserted into a result channel. In short, we have input and out put channel, and the calculation is done in the loop. So as to achieve concurrency, I have two choices, either use a go-block or use thread.
I wonder what is the intrinsic difference between these two. (We may assume there is no I/O during the calculations.) The sample code looks like the following:
(go-loop []
(when-let [input (<! input-stream)]
... ; calculations here
(>! result-chan result))
(recur))
(thread
(loop []
(when-let [input (<!! input-stream)]
... ; calculations here
(put! result-chan result))
(recur)))
I realize the number of threads that can be run simultaneously is exactly the number of CPU cores. Then in this case, is go-block and thread showing no differences if I am creating more than 8 thread or go-blocks?
I might want to simulate the differences in performance in my own laptop, but the production environment is quite different from the simulated one. I could draw no conclusions.
By the way, the calculation is not so heavy. If the inputs are not so large, 8,000 loops can be run in 1 second.
Another consideration is whether go-block vs thread will have an impact on GC performance.
There's a few things to note here.
Firstly, the thread pool that threads are created on via clojure.core.async/thread is what is known as a cached thread pool, meaning although it will re-use recently used threads inside that pool, it's essentially unbounded. Which of course means it could potentially hog a lot of system resources if left unchecked.
But given that what you're doing inside each asynchronous process is very lightweight, threads to me seem a little overkill. Of course, it's also important to take into account the quantity of items you expect to hit the input stream, if this number is large you could potentially overwhelm core.async's thread pool for go macros, potentially to the point where we're waiting for a thread to become available.
You also didn't mention preciously where you're getting the input values from, are the inputs some fixed data-set that remains constant at the start of the program, or are inputs continuously feed into the input stream from some source over time?
If it's the former then I would suggest you lean more towards transducers and I would argue that a CSP model isn't a good fit for your problem since you aren't modelling communication between separate components in your program, rather you're just processing data in parallel.
If it's the latter then I presume you have some other process that's listening to the result channel and doing something important with those results, in which case I would say your usage of go-blocks is perfectly acceptable.
Wanting to be sure we're using the correct synchronization (and no more than necessary) when writing threadsafe code in JRuby; specifically, in a Puma instantiated Rails app.
UPDATE: Extensively re-edited this question, to be very clear and use latest code we are implementing. This code uses the atomic gem written by #headius (Charles Nutter) for JRuby, but not sure it is totally necessary, or in which ways it's necessary, for what we're trying to do here.
Here's what we've got, is this overkill (meaning, are we over/uber-engineering this), or perhaps incorrect?
ourgem.rb:
require 'atomic' # gem from #headius
SUPPORTED_SERVICES = %w(serviceABC anotherSvc andSoOnSvc).freeze
module Foo
def self.included(cls)
cls.extend(ClassMethods)
cls.send :__setup
end
module ClassMethods
def get(service_name, method_name, *args)
__cached_client(service_name).send(method_name.to_sym, *args)
# we also capture exceptions here, but leaving those out for brevity
end
private
def __client(service_name)
# obtain and return a client handle for the given service_name
# we definitely want to cache the value returned from this method
# **AND**
# it is a requirement that this method ONLY be called *once PER service_name*.
end
def __cached_client(service_name)
##_clients.value[service_name]
end
def __setup
##_clients = Atomic.new({})
##_clients.update do |current_service|
SUPPORTED_SERVICES.inject(Atomic.new({}).value) do |memo, service_name|
if current_services[service_name]
current_services[service_name]
else
memo.merge({service_name => __client(service_name)})
end
end
end
end
end
end
client.rb:
require 'ourgem'
class GetStuffFromServiceABC
include Foo
def self.get_some_stuff
result = get('serviceABC', 'method_bar', 'arg1', 'arg2', 'arg3')
puts result
end
end
Summary of the above: we have ##_clients (a mutable class variable holding a Hash of clients) which we only want to populate ONCE for all available services, which are keyed on service_name.
Since the hash is in a class variable (and hence threadsafe?), are we guaranteed that the call to __client will not get run more than once per service name (even if Puma is instantiating multiple threads with this class to service all the requests from different users)? If the class variable is threadsafe (in that way), then perhaps the Atomic.new({}) is unnecessary?
Also, should we be using an Atomic.new(ThreadSafe::Hash) instead? Or again, is that not necessary?
If not (meaning: you think we do need the Atomic.news at least, and perhaps also the ThreadSafe::Hash), then why couldn't a second (or third, etc.) thread interrupt between the Atomic.new(nil) and the ##_clients.update do ... meaning the Atomic.news from EACH thread will EACH create two (separate) objects?
Thanks for any thread-safety advice, we don't see any questions on SO that directly address this issue.
Just a friendly piece of advice, before I attempt to tackle the issues you raise here:
This question, and the accompanying code, strongly suggests that you don't (yet) have a solid grasp of the issues involved in writing multi-threaded code. I encourage you to think twice before deciding to write a multi-threaded app for production use. Why do you actually want to use Puma? Is it for performance? Will your app handle many long-running, I/O-bound requests (like uploading/downloading large files) at the same time? Or (like many apps) will it primarily handle short, CPU-bound requests?
If the answer is "short/CPU-bound", then you have little to gain from using Puma. Multiple single-threaded server processes would be better. Memory consumption will be higher, but you will keep your sanity. Writing correct multi-threaded code is devilishly hard, and even experts make mistakes. If your business success, job security, etc. depends on that multi-threaded code working and working right, you are going to cause yourself a lot of unnecessary pain and mental anguish.
That aside, let me try to unravel some of the issues raised in your question. There is so much to say that it's hard to know where to start. You may want to pour yourself a cold or hot beverage of your choice before sitting down to read this treatise:
When you talk about writing "thread-safe" code, you need to be clear about what you mean. In most cases, "thread-safe" code means code which doesn't concurrently modify mutable data in a way which could cause data corruption. (What a mouthful!) That could mean that the code doesn't allow concurrent modification of mutable data at all (using locks), or that it does allow concurrent modification, but makes sure that it doesn't corrupt data (probably using atomic operations and a touch of black magic).
Note that when your threads are only reading data, not modifying it, or when working with shared stateless objects, there is no question of "thread safety".
Another definition of "thread-safe", which probably applies better to your situation, has to do with operations which affect the outside world (basically I/O). You may want some operations to only happen once, or to happen in a specific order. If the code which performs those operations runs on multiple threads, they could happen more times than desired, or in a different order than desired, unless you do something to prevent that.
It appears that your __setup method is only called when ourgem.rb is first loaded. As far as I know, even if multiple threads require the same file at the same time, MRI will only ever let a single thread load the file. I don't know whether JRuby is the same. But in any case, if your source files are being loaded more than once, that is symptomatic of a deeper problem. They should only be loaded once, on a single thread. If your app handles requests on multiple threads, those threads should be started up after the application has loaded, not before. This is the only sane way to do things.
Assuming that everything is sane, ourgem.rb will be loaded using a single thread. That means __setup will only ever be called by a single thread. In that case, there is no question of thread safety at all to worry about (as far as initialization of your "client cache" goes).
Even if __setup was to be called concurrently by multiple threads, your atomic code won't do what you think it does. First of all, you use Atomic.new({}).value. This wraps a Hash in an atomic reference, then unwraps it so you just get back the Hash. It's a no-op. You could just write {} instead.
Second, your Atomic#update call will not prevent the initialization code from running more than once. To understand this, you need to know what Atomic actually does.
Let me pull out the old, tired "increment a shared counter" example. Imagine the following code is running on 2 threads:
i += 1
We all know what can go wrong here. You may end up with the following sequence of events:
Thread A reads i and increments it.
Thread B reads i and increments it.
Thread A writes its incremented value back to i.
Thread B writes its incremented value back to i.
So we lose an update, right? But what if we store the counter value in an atomic reference, and use Atomic#update? Then it would be like this:
Thread A reads i and increments it.
Thread B reads i and increments it.
Thread A tries to write its incremented value back to i, and succeeds.
Thread B tries to write its incremented value back to i, and fails, because the value has already changed.
Thread B reads i again and increments it.
Thread B tries to write its incremented value back to i again, and succeeds this time.
Do you get the idea? Atomic never stops 2 threads from running the same code at the same time. What it does do, is force some threads to retry the #update block when necessary, to avoid lost updates.
If your goal is to ensure that your initialization code will only ever run once, using Atomic is a very inappropriate choice. If anything, it could make it run more times, rather than less (due to retries).
So, that is that. But if you're still with me here, I am actually more concerned about whether your "client" objects are themselves thread-safe. Do they have any mutable state? Since you are caching them, it seems that initializing them must be slow. Be that as it may, if you use locks to make them thread-safe, you may not be gaining anything from caching and sharing them between threads. Your "multi-threaded" server may be reduced to what is effectively an unnecessarily complicated, single-threaded server.
If the client objects have no mutable state, good for you. You can be "free and easy" and share them between threads with no problems. If they do have mutable state, but initializing them is slow, then I would recommend caching one object per thread, so they are never shared. Thread[] is your friend there.
I am using shared variables on perl with use threads::shared.
That variables can we modified only from single thread, all other threads are only 'reading' that variables.
Is it required in the 'reading' threads to lock
{
lock $shared_var;
if ($shared_var > 0) .... ;
}
?
isn't it safe to simple verification without locking (in the 'reading' thread!), like
if ($shared_var > 0) ....
?
Locking is not required to maintain internal integrity when setting or fetching a scalar.
Whether it's needed or not in your particular case depends on the needs of the reader, the other readers and the writers. It rarely makes sense not to lock, but you haven't provided enough details for us to determine what your needs are.
For example, it might not be acceptable to use an old value after the writer has updated the shared variable. For starters, this can lead to a situation where one thread is still using the old value while the another thread is using the new value, a situation that can be undesirable if those two threads interact.
It depends on whether it's meaningful to test the condition just at some point in time or other. The problem however is that in a vast majority of cases, that Boolean test means other things, which might have already changed by the time you're done reading the condition that says it represents a previous state.
Think about it. If it's an insignificant test, then it means little--and you have to question why you are making it. If it's a significant test, then it is telltale of a coherent state that may or may not exist anymore--you won't know for sure, unless you lock it.
A lot of times, say in real-time reporting, you don't really care which snapshot the database hands you, you just want a relatively current one. But, as part of its transaction logic, it keeps a complete picture of how things are prior to a commit. I don't think you're likely to find this in code, where the current state is the current state--and even a state of being in a provisional state is a definite state.
I guess one of the times this can be different is a cyclical access of a queue. If one consumer doesn't get the head record this time around, then one of them will the next time around. You can probably save some processing time, asynchronously accessing the queue counter. But here's a case where it means little in context of just one iteration.
In the case above, you would just want to put some locked-level instructions afterward that expected that the queue might actually be empty even if your test suggested it had data. So, if it is just a preliminary test, you would have to have logic that treated the test as unreliable as it actually is.
I've found myself recently using the SemaphoreSlim class to limit the work in progress of a parallelisable operation on a (large) streamed resource:
// The below code is an example of the structure of the code, there are some
// omissions around handling of tasks that do not run to completion that should be in production code
SemaphoreSlim semaphore = new SemaphoreSlim(Environment.ProcessorCount * someMagicNumber);
foreach (var result in StreamResults())
{
semaphore.Wait();
var task = DoWorkAsync(result).ContinueWith(t => semaphore.Release());
...
}
This is to avoid bringing too many results into memory and the program being unable to cope (generally evidenced via an OutOfMemoryException). Though the code works and is reasonably performant, it still feels ungainly. Notably the someMagicNumber multiplier, which although tuned via profiling, may not be as optimal as it could be and isn't resilient to changes to the implementation of DoWorkAsync.
In the same way that thread pooling can overcome the obstacle of scheduling many things for execution, I would like something that can overcome the obstacle of scheduling many things to be loaded into memory based on the resources that are available.
Since it is deterministically impossible to decide whether an OutOfMemoryException will occur, I appreciate that what I'm looking for may only be achievable via statistical means or even not at all, but I hope that I'm missing something.
Here I'd say that you're probably overthinking this problem. The consequences for overshooting are rather high (the program crashes). The consequences for being too low are that the program might be slowed down. As long as you still have some buffer beyond a minimum value, further increases to the buffer will generally have little to no effect, unless the processing time of that task in the pipe is extraordinary volatile.
If your buffer is constantly filling up it generally means that the task before it in the pipe executes quite a bit quicker than the task that follows it, so even without a fairly small buffer it is likely to always ensure the task following it has some work. The buffer size needed to get 90% of the benefits of a buffer is usually going to be quite small (a few dozen items maybe) whereas the side needed to get an OOM error are like 6+ orders of magnate higher. As long as you're somewhere in-between those two numbers (and that's a pretty big range to land in) you'll be just fine.
Just run your static tests, pick a static number, maybe add a few percent extra for "just in case" and you should be good. At most, I'd move some of the magic numbers to a config file so that they can be altered without a recompile in the event that the input data or the machine specs change radically.