Seeing by way of {hardware} counters: a journey to threefold efficiency enhance | by Netflix Expertise Weblog | Nov, 2022

By Vadim Filanovsky and Harshad Sane

In one in every of our earlier blogposts, A Microscope on Microservices we outlined three broad domains of observability (or “ranges of magnification,” as we referred to them) — Fleet-wide, Microservice and Occasion. We described the instruments and methods we use to realize perception inside every area. There’s, nevertheless, a category of issues that requires a good stronger stage of magnification going deeper down the stack to introspect CPU microarchitecture. On this blogpost we describe one such downside and the instruments we used to resolve it.

It began off as a routine migration. At Netflix, we periodically reevaluate our workloads to optimize utilization of obtainable capability. We determined to maneuver one in every of our Java microservices — let’s name it GS2 — to a bigger AWS occasion measurement, from m5.4xl (16 vCPUs) to m5.12xl (48 vCPUs). The workload of GS2 is computationally heavy the place CPU is the limiting useful resource. Whereas we perceive it’s nearly unattainable to realize a linear enhance in throughput because the variety of vCPUs develop, a near-linear enhance is attainable. Consolidating on the bigger situations reduces the amortized price of background duties, liberating up further assets for serving requests and doubtlessly offsetting the sub-linear scaling. Thus, we anticipated to roughly triple throughput per occasion from this migration, as 12xl situations have thrice the variety of vCPUs in comparison with 4xl situations. A fast canary take a look at was freed from errors and confirmed decrease latency, which is anticipated on condition that our customary canary setup routes an equal quantity of visitors to each the baseline operating on 4xl and the canary on 12xl. As GS2 depends on AWS EC2 Auto Scaling to target-track CPU utilization, we thought we simply needed to redeploy the service on the bigger occasion kind and look ahead to the ASG (Auto Scaling Group) to decide on the CPU goal. Sadly, the preliminary outcomes have been removed from our expectations:

The primary graph above represents common per-node throughput overlaid with common CPU utilization, whereas the second graph exhibits common request latency. We will see that as we reached roughly the identical CPU goal of 55%, the throughput elevated solely by ~25% on common, falling far in need of our desired purpose. What’s worse, common latency degraded by greater than 50%, with each CPU and latency patterns changing into extra “uneven.” GS2 is a stateless service that receives visitors by way of a taste of round-robin load balancer, so all nodes ought to obtain practically equal quantities of visitors. Certainly, the RPS (Requests Per Second) information exhibits little or no variation in throughput between nodes:

However as we began wanting on the breakdown of CPU and latency by node, an odd sample emerged:

Though we confirmed pretty equal visitors distribution between nodes, CPU and latency metrics surprisingly demonstrated a really totally different, bimodal distribution sample. There’s a “decrease band” of nodes exhibiting a lot decrease CPU and latency with hardly any variation; and there may be an “higher band” of nodes with considerably greater CPU/latency and huge variation. We observed solely ~12% of the nodes fall into the decrease band, a determine that was suspiciously constant over time. In each bands, efficiency traits stay constant for all the uptime of the JVM on the node, i.e. nodes by no means jumped the bands. This was our start line for troubleshooting.

Our first (and somewhat apparent) step at fixing the issue was to match flame graphs for the “gradual” and “quick” nodes. Whereas flame graphs clearly mirrored the distinction in CPU utilization because the variety of collected samples, the distribution throughout the stacks remained the identical, thus leaving us with no further perception. We turned to JVM-specific profiling, beginning with the fundamental hotspot stats, after which switching to extra detailed JFR (Java Flight Recorder) captures to match the distribution of the occasions. Once more, we got here away empty-handed as there was no noticeable distinction within the quantity or the distribution of the occasions between the “gradual” and “quick” nodes. Nonetheless suspecting one thing could be off with JIT habits, we ran some fundamental stats in opposition to image maps obtained by perf-map-agent solely to hit one other useless finish.

Satisfied we’re not lacking something on the app-, OS- and JVM- ranges, we felt the reply could be hidden at a decrease stage. Fortunately, the m5.12xl occasion kind exposes a set of core PMCs (Efficiency Monitoring Counters, a.ok.a. PMU counters), so we began by gathering a baseline set of counters utilizing PerfSpect:

Within the desk above, the nodes displaying low CPU and low latency signify a “quick node”, whereas the nodes with greater CPU/latency signify a “gradual node”. Except for apparent CPU variations, we will see that the gradual node has virtually 3x CPI (Cycles Per Instruction) of the quick node. We additionally see a lot greater L1 cache exercise mixed with 4x greater depend of MACHINE_CLEARS. One widespread trigger of those signs is so-called “false sharing” — a utilization sample occurring when 2 cores studying from / writing to unrelated variables that occur to share the identical L1 cache line. Cache line is an idea just like reminiscence web page — a contiguous chunk of knowledge (sometimes 64 bytes on x86 programs) transferred to and from the cache. This diagram illustrates it:

Every core on this diagram has its personal personal cache. Since each cores are accessing the identical reminiscence house, caches should be constant. This consistency is ensured with so-called “cache coherency protocol.” As Thread 0 writes to the “pink” variable, coherency protocol marks the entire cache line as “modified” in Thread 0’s cache and as “invalidated” in Thread 1’s cache. Later, when Thread 1 reads the “blue” variable, though the “blue” variable is just not modified, coherency protocol forces all the cache line to be reloaded from the cache that had the final modification — Thread 0’s cache on this instance. Resolving coherency throughout personal caches takes time and causes CPU stalls. Moreover, ping-ponging coherency visitors must be monitored by way of the last level shared cache’s controller, which results in much more stalls. We take CPU cache consistency without any consideration, however this “false sharing” sample illustrates there’s an enormous efficiency penalty for merely studying a variable that’s neighboring with another unrelated information.

Armed with this information, we used Intel vTune to run microarchitecture profiling. Drilling down into “sizzling” strategies and additional into the meeting code confirmed us blocks of code with some directions exceeding 100 CPI, which is extraordinarily gradual. That is the abstract of our findings:

Numbered markers from 1 to six denote the identical code/variables throughout the sources and vTune meeting view. The pink arrow signifies that the CPI worth probably belongs to the earlier instruction — that is because of the profiling skid in absence of PEBS (Processor Occasion-Based mostly Sampling), and often it’s off by a single instruction. Based mostly on the truth that (5) “repne scan” is a somewhat uncommon operation within the JVM codebase, we have been capable of hyperlink this snippet to the routine for subclass checking (the identical code exists in JDK mainline as of the writing of this blogpost). Going into the main points of subtype checking in HotSpot is much past the scope of this blogpost, however curious readers can be taught extra about it from the 2002 publication Fast Subtype Checking in the HotSpot JVM. Because of the nature of the category hierarchy used on this explicit workload, we hold hitting the code path that retains updating (6) the “_secondary_super_cache” discipline, which is a single-element cache for the last-found secondary superclass. Word how this discipline is adjoining to the “_secondary_supers”, which is a listing of all superclasses and is being learn (1) at first of the scan. A number of threads do these read-write operations, and if fields (1) and (6) fall into the identical cache line, then we hit a false sharing use case. We highlighted these fields with pink and blue colours to connect with the false sharing diagram above.

Word that because the cache line measurement is 64 bytes and the pointer measurement is 8 bytes, now we have a 1 in 8 probability of those fields falling on separate cache traces, and a 7 in 8 probability of them sharing a cache line. This 1-in-8 probability is 12.5%, matching our earlier statement on the proportion of the “quick” nodes. Fascinating!

Though the repair concerned patching the JDK, it was a easy change. We inserted padding between “_secondary_super_cache” and “_secondary_supers” fields to make sure they by no means fall into the identical cache line. Word that we didn’t change the purposeful side of JDK habits, however somewhat the info format:

The outcomes of deploying the patch have been instantly noticeable. The graph beneath is a breakdown of CPU by node. Right here we will see a red-black deployment taking place at midday, and the brand new ASG with the patched JDK taking up by 12:15:

Each CPU and latency (graph omitted for brevity) confirmed an analogous image — the “gradual” band of nodes was gone!

We didn’t have a lot time to marvel at these outcomes, nevertheless. Because the autoscaling reached our CPU goal, we observed that we nonetheless couldn’t push greater than ~150 RPS per node — effectively in need of our purpose of ~250 RPS. One other spherical of vTune profiling on the patched JDK model confirmed the identical bottleneck round secondary superclass cache lookup. It was puzzling at first to see seemingly the identical downside coming again proper after we put in a repair, however upon nearer inspection we realized we’re coping with “true sharing” now. In contrast to “false sharing,” the place 2 impartial variables share a cache line, “true sharing” refers back to the identical variable being learn and written by a number of threads/cores. On this case, CPU-enforced memory ordering is the reason for slowdown. We reasoned that eradicating the impediment of false sharing and rising the general throughput resulted in elevated execution of the identical JVM superclass caching code path. Basically, now we have greater execution concurrency, inflicting extreme strain on the superclass cache attributable to CPU-enforced reminiscence ordering protocols. The widespread method to resolve that is to keep away from writing to the shared variable altogether, successfully bypassing the JVM’s secondary superclass cache. Since this modification altered the habits of the JDK, we gated it behind a command line flag. That is the whole thing of our patch:

And listed below are the outcomes of operating with disabled superclass cache writes:

Our repair pushed the throughput to ~350 RPS on the identical CPU autoscaling goal of 55%. To place this in perspective, that’s a 3.5x enchancment over the throughput we initially reached on m5.12xl, together with a discount in each common and tail latency.

Disabling writes to the secondary superclass cache labored effectively in our case, and though this may not be a fascinating resolution in all circumstances, we wished to share our methodology, toolset and the repair within the hope that it will assist others encountering related signs. Whereas working by way of this downside, we got here throughout JDK-8180450 — a bug that’s been dormant for greater than 5 years that describes precisely the issue we have been dealing with. It appears ironic that we couldn’t discover this bug till we really discovered the reply. We consider our findings complement the nice work that has been executed in diagnosing and remediating it.

We have a tendency to think about trendy JVMs as extremely optimized runtime environments, in lots of circumstances rivaling extra “performance-oriented” languages like C++. Whereas it holds true for almost all of workloads, we have been reminded that efficiency of sure workloads operating inside JVMs may be affected not solely by the design and implementation of the applying code, but additionally by the implementation of the JVM itself. On this blogpost we described how we have been capable of leverage PMCs with a view to discover a bottleneck within the JVM’s native code, patch it, and subsequently notice higher than a threefold enhance in throughput for the workload in query. Relating to this class of efficiency points, the flexibility to introspect the execution on the stage of CPU microarchitecture proved to be the one resolution. Intel vTune offers priceless perception even with the core set of PMCs, akin to these uncovered by m5.12xl occasion kind. Exposing a extra complete set of PMCs together with PEBS throughout all occasion sorts and sizes within the cloud setting would pave the way in which for deeper efficiency evaluation and doubtlessly even bigger efficiency positive factors.
Replace: After publishing this put up we have been alerted to a separate impartial growth on this space, together with a writeup on how superclass cache affects regex pattern matching, in addition to a tool to automate the detection of JDK-8180450 utilizing an agent. Additionally of curiosity is this video describing an alternate method to diagnosing the problem. Our purpose in sharing our work is to supply info and perception to the open-source group, and it’s at all times thrilling to see (and share!) how others method related issues.