[External] : Re: jstack, profilers and other tools

[Ron Pressler]

The number of threads doesn’t “do” or not do you do anything. If requests arrive at 100K per second, each takes 500ms to process, then the number of threads you’re using *is equal to* at least 50K (assuming thread-per-request) in a stable system, that’s all. That is the physical meaning: the formula tells you what the quantities *are* in a stable system.

Because in a thread-per-request program, every concurrent request takes up at least one thread, while the formula does not immediately tell you how many machines are used, or what the RAM, CPU, and network bandwidth utilisation is, it does give you a lower bound on the total number of live threads. Conversely, the number of threads gives an upper bound on L.

As to the rest about splitting into subtasks, that increases L and reduces W by the same factor, so when applying Little’s law it’s handy to treat W as the total latency, *as if* it was processed sequentially, if we’re interested in L being the number of concurrent requests. More about that here: https://inside.java/2020/08/07/loom-performance/

— Ron

On 15 Jul 2022, at 09:37, Alex Otenko <oleksandr.otenko at gmail.com<mailto:oleksandr.otenko at gmail.com>> wrote:

You quickly jumped to a *therefore*.

Newton's second law binds force, mass and acceleration. But you can't say that you can decrease mass by increasing acceleration, if the force remains the same. That is, the statement would be arithmetically correct, but it would have no physical meaning.

Adding threads allows to do more work. But you can't do more work at will - the amount of work going through the system is a quantity independent of your design.

Now, what you could do at will, is split the work into sub-tasks. Virtual threads allow to do this at very little cost. However, you still can't talk about an increase in concurrency due to Little's law, because - enter Amdahl - response time changes.

Say, 100k requests get split into 10 sub tasks each, each runnable independently. Amdahl says your response time is going down 10-fold. So you have 100k requests times 1ms gives concurrency 100. Concurrency got reduced. Not surprising at all, because now each request spends 10x less time in the system.

What about subtasks? Aren't we running more of them? Does this mean concurrency increased?

Yes, 100k requests begets 1m sub tasks. We can't compare concurrency, because the definition of the unit of work changed: was W, became W/10. But let's see anyway. So we have 1m tasks, each finished in 1ms - concurrency is 1000. Same as before splitting the work and matching change of response time. I treat this like I would any units of measurement change.


So whereas I see a lot of good from being able to spin up threads, lots and shortlived, I don't see how you can claim concurrency increases, or that Little's law somehow controls throughput.


Alex

On Thu, 14 Jul 2022, 11:01 Ron Pressler, <ron.pressler at oracle.com<mailto:ron.pressler at oracle.com>> wrote:
Little’s law tells us what the relationship between concurrency, throughput and latency is if the system is stable. It tells us that if latency doesn’t decrease, then concurrency rises with throughput (again, if the system is stable). Therefore, to support high throughput you need a high level of concurrency. Since the Java platform’s unit of concurrency is the thread, to support high throughput you need a high number of threads. There might be other things you also need more of, but you *at least* need a high number of threads.

The number of threads is an *upper bound* on concurrency, because the platform cannot make concurrent progress on anything without a thread (with the caveat in the next paragraph). There might be other upper bounds, too (e.g. you need enough memory to concurrently store all the working data for your concurrent operations), but the number of threads *is* an upper bound, and the one virtual threads are there to remove.

Of course, as JEP 425 explains, you could abandon threads altogether and use some other construct as your unit of concurrency, but then you lose platform support.

In any event, virtual threads exist to support a high number of threads, as Little’s law requires, therefore, if you use virtual threads, you have a high number of them.

— Ron

On 14 Jul 2022, at 08:12, Alex Otenko <oleksandr.otenko at gmail.com<mailto:oleksandr.otenko at gmail.com>> wrote:

Hi Ron,

It looks you are unconvinced. Let me try with illustrative numbers.

The users opening their laptops at 9am don't know how many threads you have. So throughput remains 100k ops/sec in both setups below. Suppose, in the first setup we have a system that is stable with 1000 threads. Little's law tells us that the response time cannot exceed 10ms in this case. Little's law does not prescribe response time, by the way; it is merely a consequence of the statement that the system is stable: it couldn't have been stable if its response time were higher.

Now, let's create one thread per request. One claim is that this increases concurrency (and I object to this point alone). Suppose this means concurrency becomes 100k. Little's law says that the response time must be 1 second. Sorry, but that's hardly an improvement! In fact, for any concurrency greater than 1000 you must get response time higher than 10ms we've got with 1000 threads. This is not what we want. Fortunately, this is not what happens either.

Really, thread count in the thread per request design has little to do with concurrency level. Concurrency level is a derived quantity. It only tells us how many requests are making progress at any given time in a system that experiences request arrival rate R and which is able to process them in time T. The only thing you can control through system design is response time T.

There are good reasons to design a system that way, but Little's law is not one of them.

On Wed, 13 Jul 2022, 14:29 Ron Pressler, <ron.pressler at oracle.com<mailto:ron.pressler at oracle.com>> wrote:
The application of Little’s law is 100% correct. Little’s law tells us that the number of threads must *necessarily* rise if throughput is to be high. Whether or not that alone is *sufficient* might depend on the concurrency level of other resources as well. The number of threads is not the only quantity that limits the L in the formula, but L cannot be higher than the number of threads. Obviously, if the system’s level of concurrency is bounded at a very low level — say, 10 — then having more than 10 threads is unhelpful, but as we’re talking about a program that uses virtual threads, we know that is not the case.

Also, Little’s law describes *stable* systems; i.e. it says that *if* the system is stable, then a certain relationship must hold. While it is true that the rate of arrival might rise without bound, if the number of threads is insufficient to meet it, then the system is no longer stable (normally that means that queues are growing without bound).

— Ron

On 13 Jul 2022, at 14:00, Alex Otenko <oleksandr.otenko at gmail.com<mailto:oleksandr.otenko at gmail.com>> wrote:

This is an incorrect application of Little's Law. The law only posits that there is a connection between quantities. It doesn't specify which variables depend on which. In particular, throughput is not a free variable.

Throughput is something outside your control. 100k users open their laptops at 9am and login within 1 second - that's it, you have throughput of 100k ops/sec.

Then based on response time the system is able to deliver, you can tell what concurrency makes sense here. Adding threads is not going to change anything - certainly not if threads are not the bottleneck resource. Threads become the bottleneck when you have hardware to run them, but not the threads.

On Tue, 12 Jul 2022, 15:47 Ron Pressler, <ron.pressler at oracle.com<mailto:ron.pressler at oracle.com>> wrote:


On 11 Jul 2022, at 22:13, Rob Bygrave <robin.bygrave at gmail.com<mailto:robin.bygrave at gmail.com>> wrote:

> An existing application that migrates to using virtual threads doesn’t replace its platform threads with virtual threads

What I have been confident about to date based on the testing I've done is that we can use Jetty with a Loom based thread pool and that has worked very well. That is replacing current platform threads with virtual threads. I'm suggesting this will frequently be sub 1000 virtual threads.  Ron, are you suggesting this isn't a valid use of virtual threads or am I reading too much into what you've said here?


The throughput advantage to virtual threads comes from one aspect — their *number* — as explained by Little’s law. A web server employing virtual thread would not replace a pool of N platform threads with a pool of N virtual threads, as that does not increase the number of threads required to increase throughput. Rather, it replaces the pool of N virtual threads with an unpooled ExecutorService that spawns at least one new virtual thread for every HTTP serving task. Only that can increase the number of threads sufficiently to improve throughput.



> unusual for an application that has any virtual threads to have fewer than, say, 10,000

In the case of http server use of virtual thread, I feel the use of unusual is too strong. That is, when we are using virtual threads for application code handling of http request/response (like Jetty + Loom), I suspect this is frequently going to operate with less than 1000 concurrent requests per server instance.

1000 concurrent requests would likely translate to more than 10,000 virtual threads due to fanout (JEPs 425 and 428 cover this). In fact, even without fanout, every HTTP request might wish to spawn more than one thread, for example to have one thread for reading and one for writing. The number 10,000, however, is just illustrative. Clearly, an application with virtual threads will have some large number of threads (significantly larger than applications with just platform threads), because the ability to have a large number of threads is what virtual threads are for.

The important point is that tooling needs to adapt to a high number of threads, which is why we’ve added a tool that’s designed to make sense of many threads, where jstack might not be very useful.

— Ron




-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/loom-dev/attachments/20220715/ad6864ff/attachment-0001.htm>