Profiling Particular person Queries in a Concurrent System

CPU profiler is value its weight in gold. Measuring efficiency in-situ normally means utilizing a sampling profile. They supply a number of info whereas having very low overhead. In a concurrent system, nevertheless, it’s onerous to make use of the ensuing knowledge to extract high-level insights. Samples don’t embrace context like question IDs and application-level statistics; they present you what code was run, however not why.

This weblog introduces trampoline histories, a method Rockset has developed to effectively connect application-level info (question IDs) to the samples of a CPU profile. This lets us use profiles to grasp the efficiency of particular person queries, even when a number of queries are executing concurrently throughout the identical set of employee threads.

Primer on Rockset

Rockset is a cloud-native search and analytics database. SQL queries from a buyer are executed in a distributed trend throughout a set of servers within the cloud. We use inverted indexes, approximate vector indexes, and columnar layouts to effectively execute queries, whereas additionally processing streaming updates. The vast majority of Rockset’s performance-critical code is C++.

Most Rockset prospects have their very own devoted compute sources known as digital situations. Inside that devoted set of compute sources, nevertheless, a number of queries can execute on the similar time. Queries are executed in a distributed trend throughout all the nodes, so which means a number of queries are energetic on the similar time in the identical course of. This concurrent question execution poses a problem when attempting to measure efficiency.

Concurrent question processing improves utilization by permitting computation, I/O, and communication to be overlapped. This overlapping is very necessary for prime QPS workloads and quick queries, which have extra coordination relative to their basic work. Concurrent execution can be necessary for decreasing head-of-line blocking and latency outliers; it prevents an occasional heavy question from blocking completion of the queries that observe it.

We handle concurrency by breaking work into micro-tasks which can be run by a set set of thread swimming pools. This considerably reduces the necessity for locks, as a result of we are able to handle synchronization by way of job dependencies, and it additionally minimizes context switching overheads. Sadly, this micro-task structure makes it troublesome to profile particular person queries. Callchain samples (stack backtraces) may need come from any energetic question, so the ensuing profile exhibits solely the sum of the CPU work.

Profiles that mix all the energetic queries are higher than nothing, however a number of guide experience is required to interpret the noisy outcomes. Trampoline histories allow us to assign many of the CPU work in our execution engine to particular person question IDs, each for steady profiles and on-demand profiles. It is a very highly effective device when tuning queries or debugging anomalies.

DynamicLabel

The API we’ve constructed for including application-level metadata to the CPU samples known as DynamicLabel. Its public interface may be very easy:

class DynamicLabel {
  public:
    DynamicLabel(std::string key, std::string worth);
    ~DynamicLabel();

    template 
    std::invoke_result_t apply(Func&& func) const;
};

DynamicLabel::apply invokes func. Profile samples taken throughout that invocation could have the label hooked up.

Every question wants just one DynamicLabel. Each time a micro-task from the question is run it’s invoked by way of DynamicLabel::apply.

One of the crucial necessary properties of sampling profilers is that their overhead is proportional to their sampling price; that is what lets their overhead be made arbitrarily small. In distinction, DynamicLabel::apply should do some work for each job whatever the sampling price. In some circumstances our micro-tasks will be fairly micro, so it is necessary that apply has very low overhead.

apply‘s efficiency is the first design constraint. DynamicLabel‘s different operations (development, destruction, and label lookup throughout sampling) occur orders of magnitude much less continuously.

Let’s work by some methods we would attempt to implement the DynamicLabel performance. We’ll consider and refine them with the aim of constructing apply as quick as attainable. If you wish to skip the journey and leap straight to the vacation spot, go to the “Trampoline Histories” part.

Implementation Concepts

Concept #1: Resolve dynamic labels at pattern assortment time

The obvious option to affiliate utility metadata with a pattern is to place it there from the start. The profiler would search for dynamic labels on the similar time that it’s capturing the stack backtrace, bundling a replica of them with the callchain.

Rockset’s profiling makes use of Linux’s perf_event, the subsystem that powers the perf command line device. perf_event has many benefits over signal-based profilers (similar to gperftools). It has decrease bias, decrease skew, decrease overhead, entry to {hardware} efficiency counters, visibility into each userspace and kernel callchains, and the flexibility to measure interference from different processes. These benefits come from its structure, during which system-wide profile samples are taken by the kernel and asynchronously handed to userspace by a lock-free ring buffer.

Though perf_event has a number of benefits, we are able to’t use it for concept #1 as a result of it could possibly’t learn arbitrary userspace knowledge at sampling time. eBPF profilers have an identical limitation.

Concept #2: Report a perf pattern when the metadata modifications

If it’s not attainable to tug dynamic labels from userspace to the kernel at sampling time, then what about push? We may add an occasion to the profile each time that the thread→label mapping modifications, then post-process the profiles to match up the labels.

A method to do that could be to make use of perf uprobes. Userspace probes can file operate invocations, together with operate arguments. Sadly, uprobes are too sluggish to make use of on this trend for us. Thread pool overhead for us is about 110 nanoseconds per job. Even a single crossing from the userspace into the kernel (uprobe or syscall) would multiply this overhead.

Avoiding syscalls throughout DynamicLabel::apply additionally prevents an eBPF answer, the place we replace an eBPF map in apply after which modify an eBPF profiler like BCC to fetch the labels when sampling.

edit: eBPF can be utilized to tug from userspace when amassing a pattern, studying fsbase after which utilizing bpfprobelearnconsumer() to stroll a userspace knowledge construction that’s hooked up to a threadnative. You probably have BPF permissions enabled in your manufacturing surroundings and are utilizing a BPF-based profiler then this various could be a good one. The engineering and deployment points are extra advanced however the consequence doesn’t require in-process profile processing. Due to Jason Rahman for pointing this out.

Concept #3: Merge profiles with a userspace label historical past

If it is too costly to file modifications to the thread→label mapping within the kernel, what if we do it within the userspace? We may file a historical past of calls to DynamicLabel::apply, then be a part of it to the profile samples throughout post-processing. perf_event samples can embrace timestamps and Linux’s CLOCK_MONOTONIC clock has sufficient precision to seem strictly monotonic (not less than on the x86_64 or arm64 situations we would use), so the be a part of could be actual. A name to clock_gettime utilizing the VDSO mechanism is quite a bit sooner than a kernel transition, so the overhead could be a lot decrease than that for concept #2.

The problem with this strategy is the info footprint. DynamicLabel histories could be a number of orders of magnitude bigger than the profiles themselves, even after making use of some easy compression. Profiling is enabled repeatedly on all of our servers at a low sampling price, so attempting to persist a historical past of each micro-task invocation would rapidly overload our monitoring infrastructure.

Concept #4: In-memory historical past merging

The sooner we be a part of samples and label histories, the much less historical past we have to retailer. If we may be a part of the samples and the historical past in near-realtime (maybe each second) then we wouldn’t want to write down the histories to disk in any respect.

The commonest means to make use of Linux’s perf_event subsystem is by way of the perf command line device, however all the deep kernel magic is out there to any course of by way of the perf_event_open syscall. There are a number of configuration choices (perf_event_open(2) is the longest manpage of any system name), however when you get it arrange you possibly can learn profile samples from a lock-free ring buffer as quickly as they’re gathered by the kernel.

To keep away from rivalry, we may preserve the historical past as a set of thread-local queues that file the timestamp of each DynamicLabel::apply entry and exit. For every pattern we’d search the corresponding historical past utilizing the pattern’s timestamp.

This strategy has possible efficiency, however can we do higher?

Concept #5: Use the callchains to optimize the historical past of calls to `apply`

We will use the truth that apply exhibits up within the recorded callchains to scale back the historical past dimension. If we block inlining in order that we are able to discover DynamicLabel::apply within the name stacks, then we are able to use the backtrace to detect exit. Which means apply solely wants to write down the entry data, which file the time that an affiliation was created. Halving the variety of data halves the CPU and knowledge footprint (of the a part of the work that isn’t sampled).

This technique is the very best one but, however we are able to do even higher! The historical past entry data a spread of time for which apply was certain to a selected label, so we solely have to make a file when the binding modifications, moderately than per-invocation. This optimization will be very efficient if we’ve got a number of variations of apply to search for within the name stack. This leads us to trampoline histories, the design that we’ve got carried out and deployed.

Trampoline Histories

If the stack has sufficient info to search out the suitable DynamicLabel , then the one factor that apply must do is depart a body on the stack. Since there are a number of energetic labels, we’ll want a number of addresses.

A operate that instantly invokes one other operate is a trampoline. In C++ it would seem like this:

__attribute__((__noinline__))
void trampoline(std::move_only_function func) {
    func();
    asm risky (""); // stop tailcall optimization
}

Be aware that we have to stop compiler optimizations that will trigger the operate to not be current within the stack, specifically inlining and tailcall elimination.

The trampoline compiles to solely 5 directions, 2 to arrange the body pointer, 1 to invoke func(), and a pair of to scrub up and return. Together with padding that is 32 bytes of code.

C++ templates allow us to simply generate a complete household of trampolines, every of which has a singular tackle.

utilizing Trampoline = __attribute__((__noinline__)) void (*)(
        std::move_only_function);

constexpr size_t kNumTrampolines = ...;

template 
__attribute__((__noinline__))
void trampoline(std::move_only_function func) {
    func();
    asm risky (""); // stop tailcall optimization
}

template 
constexpr std::array makeTrampolines(
        std::index_sequence) {
    return {&trampoline...};
}

Trampoline getTrampoline(unsigned idx) {
    static constexpr auto kTrampolines =
            makeTrampolines(std::make_index_sequence{});
    return kTrampolines.at(idx);
}

We’ve now received all the low-level items we have to implement DynamicLabel:

• DynamicLabel development → discover a trampoline that isn’t presently in use, append the label and present timestamp to that trampoline’s historical past
• DynamicLabel::apply → invoke the code utilizing the trampoline
• DynamicLabel destruction → return the trampoline to a pool of unused trampolines
• Stack body symbolization → if the trampoline’s tackle is present in a callchain, search for the label within the trampoline’s historical past

Efficiency Impression

Our aim is to make DynamicLabel::apply quick, in order that we are able to use it to wrap even small items of labor. We measured it by extending our current dynamic thread pool microbenchmark, including a layer of indirection by way of apply.

{
    DynamicThreadPool executor({.maxThreads = 1});
    for (size_t i = 0; i < kNumTasks; ++i) {
        executor.add([&]() {
            label.apply([&] { ++depend; }); });
    }
    // ~DynamicThreadPool waits for all duties
}
EXPECT_EQ(kNumTasks, depend);

Maybe surprisingly, this benchmark exhibits zero efficiency impression from the additional stage of indirection, when measured utilizing both wall clock time or cycle counts. How can this be?

It seems we’re benefiting from a few years of analysis into department prediction for oblique jumps. The within of our trampoline appears like a digital technique name to the CPU. That is extraordinarily frequent, so processor distributors have put a number of effort into optimizing it.

If we use perf to measure the variety of directions within the benchmark we observe that including label.apply causes about three dozen further directions to be executed per loop. This may sluggish issues down if the CPU was front-end certain or if the vacation spot was unpredictable, however on this case we’re reminiscence certain. There are many execution sources for the additional directions, in order that they don’t truly enhance this system’s latency. Rockset is mostly reminiscence certain when executing queries; the zero-latency consequence holds in our manufacturing surroundings as nicely.

A Few Implementation Particulars

There are some things we have executed to enhance the ergonomics of our profile ecosystem:

• The perf.knowledge format emitted by perf is optimized for CPU-efficient writing, not for simplicity or ease of use. Although Rockset’s perf_event_open-based profiler pulls knowledge from perf_event_open, we’ve got chosen to emit the identical protobuf-based pprof format utilized by gperftools. Importantly, the pprof format helps arbitrary labels on samples and the pprof visualizer already has the flexibility to filter on these tags, so it was straightforward so as to add and use the data from DynamicLabel.
• We subtract one from most callchain addresses earlier than symbolizing, as a result of the return tackle is definitely the primary instruction that shall be run after returning. That is particularly necessary when utilizing inline frames, since neighboring directions are sometimes not from the identical supply operate.
• We rewrite trampoline to trampoline<0> in order that we’ve got the choice of ignoring the tags and rendering an everyday flame graph.
• When simplifying demangled constructor names, we use one thing like Foo::copy_construct and Foo::move_construct moderately than simplifying each to Foo::Foo. Differentiating constructor sorts makes it a lot simpler to seek for pointless copies. (In case you implement this be sure to can deal with demangled names with unbalanced < and >, similar to std::enable_if 4, void>::kind.)
• We compile with -fno-omit-frame-pointer and use body tips to construct our callchains, however some necessary glibc capabilities like memcpy are written in meeting and don’t contact the stack in any respect. For these capabilities, the backtrace captured by perf_event_open‘s PERF_SAMPLE_CALLCHAIN mode omits the operate that calls the meeting operate. We discover it through the use of PERF_SAMPLE_STACK_USER to file the highest 8 bytes of the stack, splicing it into the callchain when the leaf is in a kind of capabilities. That is a lot much less overhead than attempting to seize your complete backtrace with PERF_SAMPLE_STACK_USER.