A 40-line fix eliminated a 400x performance gap
316 points - yesterday at 11:00 PM
SourceYou can do even faster, about 8ns (almost an additional 10x improvement) by using software perf events: PERF_COUNT_SW_TASK_CLOCK is thread CPU time, it can be read through a shared page (so no syscall, see perf_event_mmap_page), and then you add the delta since the last context switch with a single rdtsc call within a seqlock.
This is not well documented unfortunately, and I'm not aware of open-source implementations of this.
EDIT: Or maybe not, I'm not sure if PERF_COUNT_SW_TASK_CLOCK allows to select only user time. The kernel can definitely do it, but I don't know if the wiring is there. However this definitely works for overall thread CPU time.
That's a brilliant trick. The setup overhead and permission requirements for perf_event might be heavy for arbitrary threads, but for long-lived threads it looks pretty awesome! Thanks for sharing!
Yes you need some lazy setup in thread-local state to use this. And short-lived threads should be avoided anyway :)
catlifeonmars
today at 7:39 AM
I guess if you need the concurrency/throughput you should use a userspace green thread implementation. I’m guessing most implementations of green threads multiplex onto long running os threads anyway
In a system with green threads, you typically want the CPU time of the fiber or tasklet rather than the carrier thread. In that case, you have to ask the scheduler, not the kernel.
Why do you need a seqlock? To make sure you're not context switched out between the read of the page value and the rdtsc?
Presumably you mean you just double check the page value after the rdtsc to make sure it hasn't changed and retry if it has?
Tbh I thought clock_gettime was a vdso based virtual syscall anyway
clock_gettime is not doing a syscall, it's using vdso.
clock_gettime() goes through the vDSO shim, but whether it avoids a syscall depends on the clock ID and (in some cases) the clock source. For thread-specific CPU user time, the vDSO shim cannot resolve the request in user space and must transit into the kernel. In this specific case, there is absolutely a syscall.
shermantanktop
today at 2:08 AM
Flamegraphs are wonderful.
Me: looks at my code. "sure, ok, looks alright."
Me: looks at the resulting flamegraph. "what the hell is this?!?!?"
I've found all kinds of crazy stuff in codebases this way. Static initializers that aren't static, one-line logger calls that trigger expensive serialization, heavy string-parsing calls that don't memoize patterns, etc. Unfortunately some of those are my fault.
I also like icicle graphs for this. They're flamegraphs, but aggregated in the reverse order. (I.e. if you have calls A->B->C and D->E->C, then both calls to C are aggregated together, rather than being stacked on top of B and E respectively. It can make it easier to see what's wrong when you have a bunch of distinct codepaths that all invoke a common library where you're spending too much time.)
Regular flamegraphs are good too, icicle graphs are just another tool in the toolbox.
So someone else linked the original flamegraph site [0] and it describes icicle graphs as "inverting the y axis" but that's not only what's happening, right? You bucket top-down the stack opposed to bottom-up, correct?
[0] https://www.brendangregg.com/flamegraphs.html
Right, what is needed is something trie-like, with the root being the most fine-grained call.
arethuza
today at 12:17 PM
I always found profiling performance critical code and experimenting with optimisations to be one of the most enjoyable parts of development - probably because of the number of surprises that I encountered ("Why on Earth is that so slow?").
I might be very wrong in every way but, string parsing and or manipulating and memoiziation... sound like a super strange combo? For the first you know you're already doing expensive allocations, but the 2nd is also not a pattern I really see apart from in JS codebases. Could you provide more context on how this actually bit you in the behind? memoizing strings seems like a complicated and error prone "welp it feels better now" territory in my mind so I'm genuinely curious.
shermantanktop
today at 4:21 AM
In Java it can be a bad toString() implementation hiding behind a + used for string assembly.
Or another great one: new instances of ObjectMapper created inside a method for a single call and then thrown away.
shermantanktop
today at 5:06 AM
To be clear this is often sloppy code that shouldn’t have been written. But in a legacy codebase this stuff can easily happen.
> but the 2nd is also not a pattern I really see apart from in JS codebases.
If you're referring to "one-line logger calls that trigger expensive serialization", it's also common in java.
I've never used flamegraphs but would like to know about them. Can you explain more? Or where should I start?
dummydummy1234
today at 9:32 AM
I would also try hotspot, it is a interactive viewer for perf graphs.
jerrinot
yesterday at 11:13 PM
Author here. After my last post about kernel bugs, I spent some time looking at how the JVM reports its own thread activity. It turns out that "What is the CPU time of this thread?" is/was a much more expensive question than it should be.
jacquesm
today at 12:10 AM
I don't think it is possible to talk about fractions of nanoseconds without having an extremely good idea of the stability and accuracy of your clock. At best I think you could claim there is some kind of reduction but it is super hard to make such claims in the absolute without doing a massive amount of prep work to ensure that the measured times themselves are indeed accurate. You could be off by a large fraction and never know the difference. So unless there is a hidden atomic clock involved somewhere in these measurements I think they should be qualified somehow.
Stability and accuracy, when applied to clocks, are generally about dynamic range, i.e. how good is the scale with which you are measuring time. So if you're talking about nanoseconds across a long time period, seconds or longer, then yeah, you probably should care about your clock. But when you're measuring nanoseconds out of a millisecond or microsecond, it really doesn't matter that much and you're going to be OK with the average crystal oscillator in a PC. (and if you're measuring a 10% difference like in the article, you're going to be fine with a mechanical clock as your reference if you can do the operation a billion times in a row).
jacquesm
today at 12:35 AM
This setup is a user space program on a machine that is not exclusively dedicated to the test running all kinds of interrupts (and other tasks) left, right and center through the software under test.
For something like this, you can just take several trials and look at the minimum observed time, which is when there will have been ~no interruptions.
https://github.com/facebook/folly/blob/main/folly/docs/Bench...
You don't actually know that for sure. You have only placed a new upper bound.
This seems like more of a philosophical argument than a practical one.
jacquesm
today at 11:46 AM
No, it is a very practical one and I'm actually surprised that you don't see it that way. Benchmarking is hard, and if you don't understand the basics then you can easily measure nonsense.
jerrinot
today at 12:42 PM
You raise a fair point about the percentiles. Those are reported as point estimates without confidence intervals and the implied precision overstates what system clock can deliver.
The mean does get proper statistical treatment (t-distribution confidence interval), but you're right that JMH doesn't compute confidence intervals for percentiles. Reporting p0.00 with three significant figures is ... optimistic.
That said I think the core finding survives this critique. The improvement shows up consistently across ~11 million samples at every percentile from p0.50 through p0.999.
Yes, I would expect the 'order of magnitude' value to be relatively close but the absolute values to be very imprecise.
Neywiny
yesterday at 11:45 PM
Did you look into the large spread on your distributions? Some of these span multiple orders of magnitude which is interesting
jerrinot
yesterday at 11:53 PM
Fair point. These were run on a standard dev workstation under load, which may account for the noise. I haven't done a deep dive into the outliers yet, but the distribution definitely warrants a more isolated look.
Very thankful for the 1liner tldr
edit : I had an afterthought about this because it ended up being a low quality comment ;
Bringing up such TLDR give a lot of value to reading content, especially on HN, as it provides way more inertia and let focus on -
reading this short form felt like that cool friend who gave you a heads up.
jerrinot
today at 12:26 AM
I was unsure whether to post it or not so I am glad you found it useful!
I have that 10-30s time window to fill when claude might be loading some stuff ; the 1 liner is exactly what fits in that window - it makes me wonder about the original idea of twitter now that I think of it - but since it's not the same kind of content I don't bother with it.It really feels like "here is the stuff, here's more about it if you want to" - really really appreciate that form and will definitely do the same format myself
abicklefitch
today at 3:21 AM
Quelle Suprise
furyofantares
today at 1:38 AM
> Flame graph image
> Click to zoom, open in a new tab for interactivity
I admit I did not expect "Open Image in New Tab" to do what it said on the tin. I guess I was aware that it was possible with SVG but I don't think I've ever seen it done and was really not expecting it.
Courtesy of Brendan Gregg and his flamegraph.pl scripts: https://github.com/brendangregg/FlameGraph
Normally, I use the generator included in async-profiler. It produces interactive HTML. But for this post, I used Brendan’s tool specifically to have a single, interactive SVG.
Author of the OpenJDK patch here.
Thanks for the write-up Jaromir :) For those interested, I explored memory overhead when reading /proc—including eBPF profiling and the history behind the poorly documented user-space ABI.
Full details in my write-up: https://norlinder.nu/posts/User-CPU-Time-JVM/
Hi Jonas, thanks for the work on OpenJDK and the post! I swear I hadn't seen your blog :) I finished my draft around Christmas and it’s been in the queue since. Great minds think alike, I guess.
edit: I just read your blog in full and I have to say I like it more than mine. You put a lot more rigor into it. I’m just peeking into things.
edit2: I linked your article from my post.
Which goes to show writing C, C++ or whatever systems language isn't automatically blazing fast, depending on what is being done.
Very interesting read.
Obviously a vdso read is going to be significantly faster than a syscall switching to the kernel, writing serialized data to a buffer, switching back to userland, and parsing that data.
higherhalf
today at 12:41 AM
clock_gettime() goes through vDSO, avoiding a context switch. It shows up on the flamegraph as well.
jerrinot
today at 12:47 AM
Only for some clocks (CLOCK_MONOTONIC, etc) and some clock sources. For VIRT/SCHED, the vDSO shim still has to invoke the actual syscall. You can't avoid the kernel transition when you need per-thread accounting.
touisteur
today at 7:48 AM
Oh for some time after its introduction, CLOCK_MONOTONIC_RAW wasn't vDSO'd and it took some time and syscall profiling ('huh, why do I see these as syscalls in perf record -e syscalls' ...) to understand what was going on.
higherhalf
today at 1:32 AM
Thanks, I really should've looked deeper than that.
no problem at all, I was confused too when I saw the profile for the first time.
If you look below the vDSO frame, there is still a syscall. I think that the vDSO implementation is missing a fast path for this particular clock id (it could be implemented though).
jerrinot
today at 12:49 AM
Exactly this.
Ono-Sendai
today at 8:34 AM
"look, I'm sorry, but the rule is simple:
if you made something 2x faster, you might have done something smart
if you made something 100x faster, you definitely just stopped doing something stupid"
https://x.com/rygorous/status/1271296834439282690
The QuestDB team are among the best doing it.
Love the people and their software.
Great blog Jaromir!
This is such a great writeup
It's kinda crazy the amount of plumbing required to get a few bits across the CPU.
It took seven years to address this concern following the initial bug report (2018). That seems like a lot, considering how instrumenting CPU time can be in the hot path for profiled code.
400x slower than 70ns is still only 28us. How often is the JVM calling this function?
It depends. If you’re doing continuous profiling, it’d make a call to get the current time at every method entry and exit, each of which could then add a context switch. In an absolute sense it appears to be small, but it could really add up.
This is what flame graphs are super helpful for, to see whether it’s really a problem or not.
Also, remember that every extra moment running instructions is a lost opportunity to put the CPU to sleep, so this has energy efficiency impact as well.
If you are doing continuous profiling, you are probably using a low overhead stack sampling profiler rather than recording every method entry and exit.
If it's calling it twice per function, that's enormously expensive and this is a major win.
28us is still solid amount of time
This is a great example of how a small change in the right place can outweigh years of incremental tuning.
I don't think I've ever seen less than 10x speedup after putting some effort into improving performance of "organic"/legacy code. It's always shocking how slow code can be before anyone complains.
burnt-resistor
today at 12:44 PM
I really wished™ there was an API/ABI for userland- and kernelland-defined individual virtual files at arbitrary locations, backed by processes and kernel modules respectively. I've tried pipes, overlays, and FUSE to no avail. It would greatly simply configuration management implementations while maintaining compatibility with the convention of plain text files, and there's often no need to have an actual file on any media or the expense of IOPS.
While I don't particularly like the IO overhead and churn consequences of real files for performance metrics, I get the 9p-like appeal of treating the virtual fs as a DBMS/API/ABI.
squirrellous
today at 5:37 AM
Does anyone knowledgeable know whether it’s possible to drastically reduce the overhead of reading from procfs? IIUC everything in it is in-memory, so there’s no real reason reading some data should take the order of 10us.
tomiezhang
today at 5:38 AM
cool