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

Ron Pressler ron.pressler at oracle.com
Tue Jul 19 11:01:56 UTC 2022

First, async APIs and lightweight user-mode threads were invented because we empirically know there are many systems where the number of threads is the first bottleneck on maximum throughput that they encounter. The purpose of async APIs/lightweight threads is to allow those systems to hit the other, higher, limits on capacity. Of course, there are systems that hit other bottlenecks first, but the number of threads limitation is known to be very common. If a theory tells you it shouldn’t be, then the it’s the theory that should be revised (hint: the portion of total latency spent waiting for I/O is commonly *very* high; if you want to use Little’s law for that calculation, note that adding concurrency for fanout increases L and reduces W by the same factor, so it's handy to add together the total time spent waiting for, say, 5-50 outgoing microservice calls as if they were done sequentially, and compare that against the total time spent composing the results; this total “wait latency” is frequently in the hundreds of milliseconds — far higher than the actual request latency — and easily two orders of magnitude higher than CPU time).

Second, the number of *threads* (as opposed to the number of concurrent operation) has, at most, a negligible impact on contention. If we’re talking about low-level memory contention, then what matters is the number of processing cores, not the number of threads (beyond the number of cores), and if we’re talking about other resources, then that contention is part of the other limits on concurrency (and so throughput) in the system, and the way it is reached — be it with many threads or one — is irrelevant. It is true that various scheduling algorithms — whether used on threads or on async constructs the relevant scheduling problems and algorithms are the same — could reduce some overhead, but we’re talking about effects that are orders of magnitude lower than what can be achieved by reducing artificial limits on concurrency, but could matter to get the very last drop of performance; I go through the calculation of the effect of scheduling overhead here: https://inside.java/2020/08/07/loom-performance/. In short, the impact of scheduling can only be high if the total amount of time spent on scheduling is significant when compared to the time spent waiting for I/O.

— Ron

On 19 Jul 2022, at 09:22, Alex Otenko <oleksandr.otenko at gmail.com<mailto:oleksandr.otenko at gmail.com>> wrote:

Thanks, that's what I was trying to get across, too.

Also, 10k threads per request doesn't mean that the concurrency is in thousands. In the thought experiment it is. In practice - ... well, if the systems are fine with dozens or even hundreds of threads, there should be no problem even doubling thread count, if it can double, or at least improve, throughput. In my experience this is not the case. There even are famous systems with self-tuning thread pool sizes, and I worked on the self-tuning algorithm. I have seen various apps and workloads that use that system and haven't seen any that would reach maximum thread count of a few hundred even on a fairly large machine. So whereas I never found anything wrong with the claim that thread count is one of the caps on throughput, I find the claim that allowing thread per request is going to improve concurrency problematic exactly because there are other caps. There surely are such workloads that are bottlenecked on thread count that can't grow into thousands, but in my practice I haven't seen a single one of this kind. If we had thousands of threads per CPU, they just need to be waiting so much that business logic must be very trivial.

For example, the thought experiment with 10k threads and 0.5s response time. If that is executed on a 1 CPU machine, each request must be spending 50 microseconds on CPU, and for the rest of time waiting for something. If it's waiting for a lock or a pool of resource, you may be better off having fewer threads (coarsening contention). So it better be some network connection, or something of the kind. So 499.95ms it is waiting on that, and does request parsing, response construction, etc in 50 microseconds. This sort of profile is not a very common pattern.

If we consider tens of CPUs for 10k threads, it starts to look far less impressive in terms of the number of threads.

That's all about concurrency and threads as a bottleneck resource. There are other important uses of threads, but those are not about increasing concurrency.

Ok, I reckon the topic got bashed to smithereens.


On Mon, 18 Jul 2022, 22:57 Ron Pressler, <ron.pressler at oracle.com<mailto:ron.pressler at oracle.com>> wrote:
“Concurrency rises with throughput”, which is just a mathematical fact, is not the same as the claim — that no one is making — that one can *raise* throughput by adding threads. However, it is the same as the claim that the *maximum* throughput might rise if the *maximum* number of threads is increased, because that’s just how dependent variables can work in mathematics, as I’ll try explaining.

There is no “more threads to get *better throughput*”, and there is no question about “applying” Little’s law. Little’s law is simply the maths that tells us how many requests are being concurrently served in some system. There is no getting around it. In a system with 10K requests/s, each taking 500ms on average, there *are* 5K concurrent requests. If the program is written in the thread-per-request style, then it *has* at least 5K threads. Now, if the rate of requests doubles to 20K req/s and the system doesn’t collapse, then then there must be at least 10K threads serving them.

Note that the increase in threads doesn’t raise the throughput, but it must accompany it. However, because concurrency rises with throughput, the *maximum* number of threads does pose an upper bound on throughput.

It is very important to understand the difference between “adding processing units could decrease latency in a data-parallel program” and “concurrency rises with throughput in a concurrent program.” In the former, the units are an independent variable, and in the latter they’re not — i.e. when the throughput is higher there are more threads, but adding threads doesn’t increase the throughput.

And yet, because this forms an *upper bound* on throughput, the ability to have more threads is a prerequisite to raising the maximum attainable throughput (with the thread-per-request style). So raising the number of threads cannot possibly increase throughput, and yet raising the maximum number of threads could increase maximum throughput (until it’s bounded by something else). That’s just how dependent variables work when talking about upper/lower bounds.

— Ron

On 18 Jul 2022, at 19:01, Alex Otenko <oleksandr.otenko at gmail.com<mailto:oleksandr.otenko at gmail.com>> wrote:

I think I have made it clear that I am not sceptical about the ability to spawn threads in large numbers, and that all I am sceptical about is the use of Little's law in the way you did. You made it look like one needs thousands of threads to get better throughput, whereas typical numbers are much more modest than that. In practice you can't heedlessly add more threads, as at some point you get response time degrading with no improvement to throughput.

On Sun, 17 Jul 2022, 10:59 Ron Pressler, <ron.pressler at oracle.com<mailto:ron.pressler at oracle.com>> wrote:
If your thread-per-request system is getting 10K req/s (on average), each request takes 500ms (on average) to handle, and this can be sustained (i.e. the system is stable), then it doesn’t matter how much CPU or RAM is consumed, how much network bandwidth you’re using, or even how many machines you have: the (average) number of threads you’re running *is* no less than 5K (and, in practice, will usually be several times that).

So it’s not that adding more threads is going to increase throughput (in fact, it won’t; having 1M threads will do nothing in this case), it’s that the number of threads is an upper bound on L (among all other upper bounds on L). Conversely, reaching a certain throughput requires some minimum number of threads.

As to how many thread-per-request systems do or would hit the OS-thread boundary before they hit others, that’s an empirical question, and I think it is well-established that there are many such systems, but if you’re sceptical and think that user-mode threads/asynchronous APIs have little impact, you can just wait and see.

— Ron

On 16 Jul 2022, at 20:30, Alex Otenko <oleksandr.otenko at gmail.com<mailto:oleksandr.otenko at gmail.com>> wrote:

That's the indisputable bit. The contentious part is that adding more threads is going to increase throughput.

Supposing that 10k threads are there, and you actually need them, you should get concurrency level 10k. Let's see what that means in practice.

If it is a 1-CPU machine, 10k requests in flight somewhere at any given time means they are waiting for 99.99% of time. Or, out of 1 second they spend 100 microseconds on CPU, and waiting for something for the rest of the time (or, out of 100ms response time, 10 microseconds on CPU - barely enough to parse REST request). This can't be the case for the majority of workflows.

Of course, having 10k threads for less than 1 second each doesn't mean you are getting concurrency thar is unattainable with fewer threads.

The bottom line is that adding threads you aren't necessarily increasing concurrency.

On Fri, 15 Jul 2022, 10:19 Ron Pressler, <ron.pressler at oracle.com<mailto:ron.pressler at oracle.com>> wrote:
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/<https://urldefense.com/v3/__https://inside.java/2020/08/07/loom-performance/__;!!ACWV5N9M2RV99hQ!LwEzCaeHJaxabByuY70NvD7RKrrcLhHjnaTLRUpLB40wh_ccQ_20JK__7-UA4AfkP2a6jEOR6KehkhXmKBfKMluHMPsa$>

— 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.


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/20220719/6a21995b/attachment-0001.htm>

More information about the loom-dev mailing list