From 5c2356988e8beea0418e1292da2adf4b988340ad Mon Sep 17 00:00:00 2001 From: Christian Schwarz Date: Mon, 25 Nov 2024 16:52:39 +0100 Subject: [PATCH] page_service: add benchmark for batching (#9820) This PR adds two benchmark to demonstrate the effect of server-side getpage request batching added in https://github.com/neondatabase/neon/pull/9321. For the CPU usage, I found the the `prometheus` crate's built-in CPU usage accounts the seconds at integer granularity. That's not enough you reduce the target benchmark runtime for local iteration. So, add a new `libmetrics` metric and report that. The benchmarks are disabled because [on our benchmark nodes, timer resolution isn't high enough](https://neondb.slack.com/archives/C059ZC138NR/p1732264223207449). They work (no statement about quality) on my bare-metal devbox. They will be refined and enabled once we find a fix. Candidates at time of writing are: - https://github.com/neondatabase/neon/pull/9822 - https://github.com/neondatabase/neon/pull/9851 Refs: - Epic: https://github.com/neondatabase/neon/issues/9376 - Extracted from https://github.com/neondatabase/neon/pull/9792 --- libs/metrics/src/more_process_metrics.rs | 40 ++- test_runner/performance/README.md | 3 +- .../test_pageserver_getpage_merge.py | 307 ++++++++++++++++++ 3 files changed, 347 insertions(+), 3 deletions(-) create mode 100644 test_runner/performance/pageserver/test_pageserver_getpage_merge.py diff --git a/libs/metrics/src/more_process_metrics.rs b/libs/metrics/src/more_process_metrics.rs index 920724fdecb4..13a745e03148 100644 --- a/libs/metrics/src/more_process_metrics.rs +++ b/libs/metrics/src/more_process_metrics.rs @@ -2,14 +2,28 @@ // This module has heavy inspiration from the prometheus crate's `process_collector.rs`. +use once_cell::sync::Lazy; +use prometheus::Gauge; + use crate::UIntGauge; pub struct Collector { descs: Vec, vmlck: crate::UIntGauge, + cpu_seconds_highres: Gauge, } -const NMETRICS: usize = 1; +const NMETRICS: usize = 2; + +static CLK_TCK_F64: Lazy = Lazy::new(|| { + let long = unsafe { libc::sysconf(libc::_SC_CLK_TCK) }; + if long == -1 { + panic!("sysconf(_SC_CLK_TCK) failed"); + } + let convertible_to_f64: i32 = + i32::try_from(long).expect("sysconf(_SC_CLK_TCK) is larger than i32"); + convertible_to_f64 as f64 +}); impl prometheus::core::Collector for Collector { fn desc(&self) -> Vec<&prometheus::core::Desc> { @@ -27,6 +41,12 @@ impl prometheus::core::Collector for Collector { mfs.extend(self.vmlck.collect()) } } + if let Ok(stat) = myself.stat() { + let cpu_seconds = stat.utime + stat.stime; + self.cpu_seconds_highres + .set(cpu_seconds as f64 / *CLK_TCK_F64); + mfs.extend(self.cpu_seconds_highres.collect()); + } mfs } } @@ -43,7 +63,23 @@ impl Collector { .cloned(), ); - Self { descs, vmlck } + let cpu_seconds_highres = Gauge::new( + "libmetrics_process_cpu_seconds_highres", + "Total user and system CPU time spent in seconds.\ + Sub-second resolution, hence better than `process_cpu_seconds_total`.", + ) + .unwrap(); + descs.extend( + prometheus::core::Collector::desc(&cpu_seconds_highres) + .into_iter() + .cloned(), + ); + + Self { + descs, + vmlck, + cpu_seconds_highres, + } } } diff --git a/test_runner/performance/README.md b/test_runner/performance/README.md index 70d75a6dcfa9..85096d377004 100644 --- a/test_runner/performance/README.md +++ b/test_runner/performance/README.md @@ -15,6 +15,7 @@ Some handy pytest flags for local development: - `-k` selects a test to run - `--timeout=0` disables our default timeout of 300s (see `setup.cfg`) - `--preserve-database-files` to skip cleanup +- `--out-dir` to produce a JSON with the recorded test metrics # What performance tests do we have and how we run them @@ -36,6 +37,6 @@ All tests run only once. Usually to obtain more consistent performance numbers, ## Results collection -Local test results for main branch, and results of daily performance tests, are stored in a neon project deployed in production environment. There is a Grafana dashboard that visualizes the results. Here is the [dashboard](https://observer.zenith.tech/d/DGKBm9Jnz/perf-test-results?orgId=1). The main problem with it is the unavailability to point at particular commit, though the data for that is available in the database. Needs some tweaking from someone who knows Grafana tricks. +Local test results for main branch, and results of daily performance tests, are stored in a [neon project](https://console.neon.tech/app/projects/withered-sky-69117821) deployed in production environment. There is a Grafana dashboard that visualizes the results. Here is the [dashboard](https://observer.zenith.tech/d/DGKBm9Jnz/perf-test-results?orgId=1). The main problem with it is the unavailability to point at particular commit, though the data for that is available in the database. Needs some tweaking from someone who knows Grafana tricks. There is also an inconsistency in test naming. Test name should be the same across platforms, and results can be differentiated by the platform field. But currently, platform is sometimes included in test name because of the way how parametrization works in pytest. I.e. there is a platform switch in the dashboard with neon-local-ci and neon-staging variants. I.e. some tests under neon-local-ci value for a platform switch are displayed as `Test test_runner/performance/test_bulk_insert.py::test_bulk_insert[vanilla]` and `Test test_runner/performance/test_bulk_insert.py::test_bulk_insert[neon]` which is highly confusing. diff --git a/test_runner/performance/pageserver/test_pageserver_getpage_merge.py b/test_runner/performance/pageserver/test_pageserver_getpage_merge.py new file mode 100644 index 000000000000..34cce9900b1f --- /dev/null +++ b/test_runner/performance/pageserver/test_pageserver_getpage_merge.py @@ -0,0 +1,307 @@ +import dataclasses +import json +import time +from dataclasses import dataclass +from pathlib import Path +from typing import Any + +import pytest +from fixtures.benchmark_fixture import MetricReport, NeonBenchmarker +from fixtures.log_helper import log +from fixtures.neon_fixtures import NeonEnvBuilder, PgBin, wait_for_last_flush_lsn +from fixtures.utils import humantime_to_ms + +TARGET_RUNTIME = 60 + + +@pytest.mark.skip("See https://github.com/neondatabase/neon/pull/9820#issue-2675856095") +@pytest.mark.parametrize( + "tablesize_mib, batch_timeout, target_runtime, effective_io_concurrency, readhead_buffer_size, name", + [ + # the next 4 cases demonstrate how not-batchable workloads suffer from batching timeout + (50, None, TARGET_RUNTIME, 1, 128, "not batchable no batching"), + (50, "10us", TARGET_RUNTIME, 1, 128, "not batchable 10us timeout"), + (50, "1ms", TARGET_RUNTIME, 1, 128, "not batchable 1ms timeout"), + # the next 4 cases demonstrate how batchable workloads benefit from batching + (50, None, TARGET_RUNTIME, 100, 128, "batchable no batching"), + (50, "10us", TARGET_RUNTIME, 100, 128, "batchable 10us timeout"), + (50, "100us", TARGET_RUNTIME, 100, 128, "batchable 100us timeout"), + (50, "1ms", TARGET_RUNTIME, 100, 128, "batchable 1ms timeout"), + ], +) +def test_getpage_merge_smoke( + neon_env_builder: NeonEnvBuilder, + zenbenchmark: NeonBenchmarker, + tablesize_mib: int, + batch_timeout: str | None, + target_runtime: int, + effective_io_concurrency: int, + readhead_buffer_size: int, + name: str, +): + """ + Do a bunch of sequential scans and ensure that the pageserver does some merging. + """ + + # + # record perf-related parameters as metrics to simplify processing of results + # + params: dict[str, tuple[float | int, dict[str, Any]]] = {} + + params.update( + { + "tablesize_mib": (tablesize_mib, {"unit": "MiB"}), + "batch_timeout": ( + -1 if batch_timeout is None else 1e3 * humantime_to_ms(batch_timeout), + {"unit": "us"}, + ), + # target_runtime is just a polite ask to the workload to run for this long + "effective_io_concurrency": (effective_io_concurrency, {}), + "readhead_buffer_size": (readhead_buffer_size, {}), + # name is not a metric + } + ) + + log.info("params: %s", params) + + for param, (value, kwargs) in params.items(): + zenbenchmark.record( + param, + metric_value=value, + unit=kwargs.pop("unit", ""), + report=MetricReport.TEST_PARAM, + **kwargs, + ) + + # + # Setup + # + + env = neon_env_builder.init_start() + ps_http = env.pageserver.http_client() + endpoint = env.endpoints.create_start("main") + conn = endpoint.connect() + cur = conn.cursor() + + cur.execute("SET max_parallel_workers_per_gather=0") # disable parallel backends + cur.execute(f"SET effective_io_concurrency={effective_io_concurrency}") + cur.execute( + f"SET neon.readahead_buffer_size={readhead_buffer_size}" + ) # this is the current default value, but let's hard-code that + + cur.execute("CREATE EXTENSION IF NOT EXISTS neon;") + cur.execute("CREATE EXTENSION IF NOT EXISTS neon_test_utils;") + + log.info("Filling the table") + cur.execute("CREATE TABLE t (data char(1000)) with (fillfactor=10)") + tablesize = tablesize_mib * 1024 * 1024 + npages = tablesize // (8 * 1024) + cur.execute("INSERT INTO t SELECT generate_series(1, %s)", (npages,)) + # TODO: can we force postgres to do sequential scans? + + # + # Run the workload, collect `Metrics` before and after, calculate difference, normalize. + # + + @dataclass + class Metrics: + time: float + pageserver_getpage_count: float + pageserver_vectored_get_count: float + compute_getpage_count: float + pageserver_cpu_seconds_total: float + + def __sub__(self, other: "Metrics") -> "Metrics": + return Metrics( + time=self.time - other.time, + pageserver_getpage_count=self.pageserver_getpage_count + - other.pageserver_getpage_count, + pageserver_vectored_get_count=self.pageserver_vectored_get_count + - other.pageserver_vectored_get_count, + compute_getpage_count=self.compute_getpage_count - other.compute_getpage_count, + pageserver_cpu_seconds_total=self.pageserver_cpu_seconds_total + - other.pageserver_cpu_seconds_total, + ) + + def normalize(self, by) -> "Metrics": + return Metrics( + time=self.time / by, + pageserver_getpage_count=self.pageserver_getpage_count / by, + pageserver_vectored_get_count=self.pageserver_vectored_get_count / by, + compute_getpage_count=self.compute_getpage_count / by, + pageserver_cpu_seconds_total=self.pageserver_cpu_seconds_total / by, + ) + + def get_metrics() -> Metrics: + with conn.cursor() as cur: + cur.execute( + "select value from neon_perf_counters where metric='getpage_wait_seconds_count';" + ) + compute_getpage_count = cur.fetchall()[0][0] + pageserver_metrics = ps_http.get_metrics() + return Metrics( + time=time.time(), + pageserver_getpage_count=pageserver_metrics.query_one( + "pageserver_smgr_query_seconds_count", {"smgr_query_type": "get_page_at_lsn"} + ).value, + pageserver_vectored_get_count=pageserver_metrics.query_one( + "pageserver_get_vectored_seconds_count", {"task_kind": "PageRequestHandler"} + ).value, + compute_getpage_count=compute_getpage_count, + pageserver_cpu_seconds_total=pageserver_metrics.query_one( + "libmetrics_process_cpu_seconds_highres" + ).value, + ) + + def workload() -> Metrics: + start = time.time() + iters = 0 + while time.time() - start < target_runtime or iters < 2: + log.info("Seqscan %d", iters) + if iters == 1: + # round zero for warming up + before = get_metrics() + cur.execute( + "select clear_buffer_cache()" + ) # TODO: what about LFC? doesn't matter right now because LFC isn't enabled by default in tests + cur.execute("select sum(data::bigint) from t") + assert cur.fetchall()[0][0] == npages * (npages + 1) // 2 + iters += 1 + after = get_metrics() + return (after - before).normalize(iters - 1) + + env.pageserver.patch_config_toml_nonrecursive({"server_side_batch_timeout": batch_timeout}) + env.pageserver.restart() + metrics = workload() + + log.info("Results: %s", metrics) + + # + # Sanity-checks on the collected data + # + # assert that getpage counts roughly match between compute and ps + assert metrics.pageserver_getpage_count == pytest.approx( + metrics.compute_getpage_count, rel=0.01 + ) + + # + # Record the results + # + + for metric, value in dataclasses.asdict(metrics).items(): + zenbenchmark.record(f"counters.{metric}", value, unit="", report=MetricReport.TEST_PARAM) + + zenbenchmark.record( + "perfmetric.batching_factor", + metrics.pageserver_getpage_count / metrics.pageserver_vectored_get_count, + unit="", + report=MetricReport.HIGHER_IS_BETTER, + ) + + +@pytest.mark.skip("See https://github.com/neondatabase/neon/pull/9820#issue-2675856095") +@pytest.mark.parametrize( + "batch_timeout", [None, "10us", "20us", "50us", "100us", "200us", "500us", "1ms"] +) +def test_timer_precision( + neon_env_builder: NeonEnvBuilder, + zenbenchmark: NeonBenchmarker, + pg_bin: PgBin, + batch_timeout: str | None, +): + """ + Determine the batching timeout precision (mean latency) and tail latency impact. + + The baseline is `None`; an ideal batching timeout implementation would increase + the mean latency by exactly `batch_timeout`. + + That is not the case with the current implementation, will be addressed in future changes. + """ + + # + # Setup + # + + def patch_ps_config(ps_config): + ps_config["server_side_batch_timeout"] = batch_timeout + + neon_env_builder.pageserver_config_override = patch_ps_config + + env = neon_env_builder.init_start() + endpoint = env.endpoints.create_start("main") + conn = endpoint.connect() + cur = conn.cursor() + + cur.execute("SET max_parallel_workers_per_gather=0") # disable parallel backends + cur.execute("SET effective_io_concurrency=1") + + cur.execute("CREATE EXTENSION IF NOT EXISTS neon;") + cur.execute("CREATE EXTENSION IF NOT EXISTS neon_test_utils;") + + log.info("Filling the table") + cur.execute("CREATE TABLE t (data char(1000)) with (fillfactor=10)") + tablesize = 50 * 1024 * 1024 + npages = tablesize // (8 * 1024) + cur.execute("INSERT INTO t SELECT generate_series(1, %s)", (npages,)) + # TODO: can we force postgres to do sequential scans? + + cur.close() + conn.close() + + wait_for_last_flush_lsn(env, endpoint, env.initial_tenant, env.initial_timeline) + + endpoint.stop() + + for sk in env.safekeepers: + sk.stop() + + # + # Run single-threaded pagebench (TODO: dedup with other benchmark code) + # + + env.pageserver.allowed_errors.append( + # https://github.com/neondatabase/neon/issues/6925 + r".*query handler for.*pagestream.*failed: unexpected message: CopyFail during COPY.*" + ) + + ps_http = env.pageserver.http_client() + + cmd = [ + str(env.neon_binpath / "pagebench"), + "get-page-latest-lsn", + "--mgmt-api-endpoint", + ps_http.base_url, + "--page-service-connstring", + env.pageserver.connstr(password=None), + "--num-clients", + "1", + "--runtime", + "10s", + ] + log.info(f"command: {' '.join(cmd)}") + basepath = pg_bin.run_capture(cmd, with_command_header=False) + results_path = Path(basepath + ".stdout") + log.info(f"Benchmark results at: {results_path}") + + with open(results_path) as f: + results = json.load(f) + log.info(f"Results:\n{json.dumps(results, sort_keys=True, indent=2)}") + + total = results["total"] + + metric = "latency_mean" + zenbenchmark.record( + metric, + metric_value=humantime_to_ms(total[metric]), + unit="ms", + report=MetricReport.LOWER_IS_BETTER, + ) + + metric = "latency_percentiles" + for k, v in total[metric].items(): + zenbenchmark.record( + f"{metric}.{k}", + metric_value=humantime_to_ms(v), + unit="ms", + report=MetricReport.LOWER_IS_BETTER, + )