Line data Source code
1 : use std::collections::{HashMap, HashSet, VecDeque};
2 : use std::future::Future;
3 : use std::num::NonZeroUsize;
4 : use std::pin::Pin;
5 : use std::sync::atomic::{AtomicU64, Ordering};
6 : use std::sync::{Arc, Mutex};
7 : use std::time::{Duration, Instant};
8 :
9 : use anyhow::Context;
10 : use async_trait::async_trait;
11 : use bytes::Bytes;
12 : use camino::Utf8PathBuf;
13 : use futures::{Stream, StreamExt as _};
14 : use pageserver_api::key::Key;
15 : use pageserver_api::keyspace::KeySpaceAccum;
16 : use pageserver_api::pagestream_api::{PagestreamGetPageRequest, PagestreamRequest};
17 : use pageserver_api::reltag::RelTag;
18 : use pageserver_api::shard::TenantShardId;
19 : use pageserver_page_api as page_api;
20 : use rand::prelude::*;
21 : use tokio::task::JoinSet;
22 : use tokio_util::sync::CancellationToken;
23 : use tracing::info;
24 : use url::Url;
25 : use utils::id::TenantTimelineId;
26 : use utils::lsn::Lsn;
27 : use utils::shard::ShardIndex;
28 :
29 : use crate::util::tokio_thread_local_stats::AllThreadLocalStats;
30 : use crate::util::{request_stats, tokio_thread_local_stats};
31 :
32 : /// GetPage@LatestLSN, uniformly distributed across the compute-accessible keyspace.
33 : #[derive(clap::Parser)]
34 : pub(crate) struct Args {
35 : #[clap(long, default_value = "http://localhost:9898")]
36 : mgmt_api_endpoint: String,
37 : /// Pageserver connection string. Supports postgresql:// and grpc:// protocols.
38 : #[clap(long, default_value = "postgres://postgres@localhost:64000")]
39 : page_service_connstring: String,
40 : #[clap(long)]
41 : pageserver_jwt: Option<String>,
42 : #[clap(long, default_value = "1")]
43 : num_clients: NonZeroUsize,
44 : #[clap(long)]
45 : runtime: Option<humantime::Duration>,
46 : /// If true, enable compression (only for gRPC).
47 : #[clap(long)]
48 : compression: bool,
49 : /// Each client sends requests at the given rate.
50 : ///
51 : /// If a request takes too long and we should be issuing a new request already,
52 : /// we skip that request and account it as `MISSED`.
53 : #[clap(long)]
54 : per_client_rate: Option<usize>,
55 : /// Probability for sending `latest=true` in the request (uniform distribution).
56 : #[clap(long, default_value = "1")]
57 : req_latest_probability: f64,
58 : #[clap(long)]
59 : limit_to_first_n_targets: Option<usize>,
60 : /// For large pageserver installations, enumerating the keyspace takes a lot of time.
61 : /// If specified, the specified path is used to maintain a cache of the keyspace enumeration result.
62 : /// The cache is tagged and auto-invalided by the tenant/timeline ids only.
63 : /// It doesn't get invalidated if the keyspace changes under the hood, e.g., due to new ingested data or compaction.
64 : #[clap(long)]
65 : keyspace_cache: Option<Utf8PathBuf>,
66 : /// Before starting the benchmark, live-reconfigure the pageserver to use the given
67 : /// [`pageserver_api::models::virtual_file::IoEngineKind`].
68 : #[clap(long)]
69 : set_io_engine: Option<pageserver_api::models::virtual_file::IoEngineKind>,
70 :
71 : /// Before starting the benchmark, live-reconfigure the pageserver to use specified io mode (buffered vs. direct).
72 : #[clap(long)]
73 : set_io_mode: Option<pageserver_api::models::virtual_file::IoMode>,
74 :
75 : /// Queue depth generated in each client.
76 : #[clap(long, default_value = "1")]
77 : queue_depth: NonZeroUsize,
78 :
79 : /// Batch size of contiguous pages generated by each client. This is equivalent to how Postgres
80 : /// will request page batches (e.g. prefetches or vectored reads). A batch counts as 1 RPS and
81 : /// 1 queue depth.
82 : ///
83 : /// The libpq protocol does not support client-side batching, and will submit batches as many
84 : /// individual requests, in the hope that the server will batch them. Each batch still counts as
85 : /// 1 RPS and 1 queue depth.
86 : #[clap(long, default_value = "1")]
87 : batch_size: NonZeroUsize,
88 :
89 : #[clap(long)]
90 : only_relnode: Option<u32>,
91 :
92 : targets: Option<Vec<TenantTimelineId>>,
93 : }
94 :
95 : /// State shared by all clients
96 : #[derive(Debug)]
97 : struct SharedState {
98 : start_work_barrier: tokio::sync::Barrier,
99 : live_stats: LiveStats,
100 : }
101 :
102 : #[derive(Debug, Default)]
103 : struct LiveStats {
104 : completed_requests: AtomicU64,
105 : missed: AtomicU64,
106 : }
107 :
108 : impl LiveStats {
109 0 : fn request_done(&self) {
110 0 : self.completed_requests.fetch_add(1, Ordering::Relaxed);
111 0 : }
112 0 : fn missed(&self, n: u64) {
113 0 : self.missed.fetch_add(n, Ordering::Relaxed);
114 0 : }
115 : }
116 :
117 0 : #[derive(Clone, serde::Serialize, serde::Deserialize)]
118 : struct KeyRange {
119 : timeline: TenantTimelineId,
120 : timeline_lsn: Lsn,
121 : start: i128,
122 : end: i128,
123 : }
124 :
125 : impl KeyRange {
126 0 : fn len(&self) -> i128 {
127 0 : self.end - self.start
128 0 : }
129 : }
130 :
131 : #[derive(PartialEq, Eq, Hash, Copy, Clone)]
132 : struct WorkerId {
133 : timeline: TenantTimelineId,
134 : num_client: usize, // from 0..args.num_clients
135 : }
136 :
137 : #[derive(serde::Serialize)]
138 : struct Output {
139 : total: request_stats::Output,
140 : }
141 :
142 : tokio_thread_local_stats::declare!(STATS: request_stats::Stats);
143 :
144 0 : pub(crate) fn main(args: Args) -> anyhow::Result<()> {
145 0 : tokio_thread_local_stats::main!(STATS, move |thread_local_stats| {
146 0 : main_impl(args, thread_local_stats)
147 0 : })
148 0 : }
149 :
150 0 : async fn main_impl(
151 0 : args: Args,
152 0 : all_thread_local_stats: AllThreadLocalStats<request_stats::Stats>,
153 0 : ) -> anyhow::Result<()> {
154 0 : let args: &'static Args = Box::leak(Box::new(args));
155 :
156 0 : let mgmt_api_client = Arc::new(pageserver_client::mgmt_api::Client::new(
157 0 : reqwest::Client::new(), // TODO: support ssl_ca_file for https APIs in pagebench.
158 0 : args.mgmt_api_endpoint.clone(),
159 0 : args.pageserver_jwt.as_deref(),
160 : ));
161 :
162 0 : if let Some(engine_str) = &args.set_io_engine {
163 0 : mgmt_api_client.put_io_engine(engine_str).await?;
164 0 : }
165 :
166 0 : if let Some(mode) = &args.set_io_mode {
167 0 : mgmt_api_client.put_io_mode(mode).await?;
168 0 : }
169 :
170 : // discover targets
171 0 : let timelines: Vec<TenantTimelineId> = crate::util::cli::targets::discover(
172 0 : &mgmt_api_client,
173 0 : crate::util::cli::targets::Spec {
174 0 : limit_to_first_n_targets: args.limit_to_first_n_targets,
175 0 : targets: args.targets.clone(),
176 0 : },
177 0 : )
178 0 : .await?;
179 :
180 0 : #[derive(serde::Deserialize)]
181 : struct KeyspaceCacheDe {
182 : tag: Vec<TenantTimelineId>,
183 : data: Vec<KeyRange>,
184 : }
185 : #[derive(serde::Serialize)]
186 : struct KeyspaceCacheSer<'a> {
187 : tag: &'a [TenantTimelineId],
188 : data: &'a [KeyRange],
189 : }
190 0 : let cache = args
191 0 : .keyspace_cache
192 0 : .as_ref()
193 0 : .map(|keyspace_cache_file| {
194 0 : let contents = match std::fs::read(keyspace_cache_file) {
195 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
196 0 : return anyhow::Ok(None);
197 : }
198 0 : x => x.context("read keyspace cache file")?,
199 : };
200 0 : let cache: KeyspaceCacheDe =
201 0 : serde_json::from_slice(&contents).context("deserialize cache file")?;
202 0 : let tag_ok = HashSet::<TenantTimelineId>::from_iter(cache.tag.into_iter())
203 0 : == HashSet::from_iter(timelines.iter().cloned());
204 0 : info!("keyspace cache file matches tag: {tag_ok}");
205 0 : anyhow::Ok(if tag_ok { Some(cache.data) } else { None })
206 0 : })
207 0 : .transpose()?
208 0 : .flatten();
209 0 : let all_ranges: Vec<KeyRange> = if let Some(cached) = cache {
210 0 : info!("using keyspace cache file");
211 0 : cached
212 : } else {
213 0 : let mut js = JoinSet::new();
214 0 : for timeline in &timelines {
215 0 : js.spawn({
216 0 : let mgmt_api_client = Arc::clone(&mgmt_api_client);
217 0 : let timeline = *timeline;
218 0 : async move {
219 0 : let partitioning = mgmt_api_client
220 0 : .keyspace(
221 0 : TenantShardId::unsharded(timeline.tenant_id),
222 0 : timeline.timeline_id,
223 0 : )
224 0 : .await?;
225 0 : let lsn = partitioning.at_lsn;
226 0 : let start = Instant::now();
227 0 : let mut filtered = KeySpaceAccum::new();
228 : // let's hope this is inlined and vectorized...
229 : // TODO: turn this loop into a is_rel_block_range() function.
230 0 : for r in partitioning.keys.ranges.iter() {
231 0 : let mut i = r.start;
232 0 : while i != r.end {
233 0 : let mut include = true;
234 0 : include &= i.is_rel_block_key();
235 0 : if let Some(only_relnode) = args.only_relnode {
236 0 : include &= i.is_rel_block_of_rel(only_relnode);
237 0 : }
238 0 : if include {
239 0 : filtered.add_key(i);
240 0 : }
241 0 : i = i.next();
242 : }
243 : }
244 0 : let filtered = filtered.to_keyspace();
245 0 : let filter_duration = start.elapsed();
246 :
247 0 : anyhow::Ok((
248 0 : filter_duration,
249 0 : filtered.ranges.into_iter().map(move |r| KeyRange {
250 0 : timeline,
251 0 : timeline_lsn: lsn,
252 0 : start: r.start.to_i128(),
253 0 : end: r.end.to_i128(),
254 0 : }),
255 : ))
256 0 : }
257 : });
258 : }
259 0 : let mut total_filter_duration = Duration::from_secs(0);
260 0 : let mut all_ranges: Vec<KeyRange> = Vec::new();
261 0 : while let Some(res) = js.join_next().await {
262 0 : let (filter_duration, range) = res.unwrap().unwrap();
263 0 : all_ranges.extend(range);
264 0 : total_filter_duration += filter_duration;
265 0 : }
266 0 : info!("filter duration: {}", total_filter_duration.as_secs_f64());
267 0 : if let Some(cachefile) = args.keyspace_cache.as_ref() {
268 0 : let cache = KeyspaceCacheSer {
269 0 : tag: &timelines,
270 0 : data: &all_ranges,
271 0 : };
272 0 : let bytes = serde_json::to_vec(&cache).context("serialize keyspace for cache file")?;
273 0 : std::fs::write(cachefile, bytes).context("write keyspace cache file to disk")?;
274 0 : info!("successfully wrote keyspace cache file");
275 0 : }
276 0 : all_ranges
277 : };
278 :
279 0 : let num_live_stats_dump = 1;
280 0 : let num_work_sender_tasks = args.num_clients.get() * timelines.len();
281 0 : let num_main_impl = 1;
282 :
283 0 : let shared_state = Arc::new(SharedState {
284 0 : start_work_barrier: tokio::sync::Barrier::new(
285 0 : num_live_stats_dump + num_work_sender_tasks + num_main_impl,
286 0 : ),
287 0 : live_stats: LiveStats::default(),
288 0 : });
289 0 : let cancel = CancellationToken::new();
290 :
291 0 : let ss = shared_state.clone();
292 0 : tokio::spawn({
293 0 : async move {
294 0 : ss.start_work_barrier.wait().await;
295 : loop {
296 0 : let start = std::time::Instant::now();
297 0 : tokio::time::sleep(std::time::Duration::from_secs(1)).await;
298 0 : let stats = &ss.live_stats;
299 0 : let completed_requests = stats.completed_requests.swap(0, Ordering::Relaxed);
300 0 : let missed = stats.missed.swap(0, Ordering::Relaxed);
301 0 : let elapsed = start.elapsed();
302 0 : info!(
303 0 : "RPS: {:.0} MISSED: {:.0}",
304 0 : completed_requests as f64 / elapsed.as_secs_f64(),
305 0 : missed as f64 / elapsed.as_secs_f64()
306 : );
307 : }
308 : }
309 : });
310 :
311 0 : let rps_period = args
312 0 : .per_client_rate
313 0 : .map(|rps_limit| Duration::from_secs_f64(1.0 / (rps_limit as f64)));
314 0 : let make_worker: &dyn Fn(WorkerId) -> Pin<Box<dyn Send + Future<Output = ()>>> = &|worker_id| {
315 0 : let ss = shared_state.clone();
316 0 : let cancel = cancel.clone();
317 0 : let ranges: Vec<KeyRange> = all_ranges
318 0 : .iter()
319 0 : .filter(|r| r.timeline == worker_id.timeline)
320 0 : .cloned()
321 0 : .collect();
322 0 : let weights =
323 0 : rand::distributions::weighted::WeightedIndex::new(ranges.iter().map(|v| v.len()))
324 0 : .unwrap();
325 :
326 0 : Box::pin(async move {
327 0 : let scheme = match Url::parse(&args.page_service_connstring) {
328 0 : Ok(url) => url.scheme().to_lowercase().to_string(),
329 0 : Err(url::ParseError::RelativeUrlWithoutBase) => "postgresql".to_string(),
330 0 : Err(err) => panic!("invalid connstring: {err}"),
331 : };
332 0 : let client: Box<dyn Client> = match scheme.as_str() {
333 0 : "postgresql" | "postgres" => {
334 0 : assert!(!args.compression, "libpq does not support compression");
335 0 : Box::new(
336 0 : LibpqClient::new(&args.page_service_connstring, worker_id.timeline)
337 0 : .await
338 0 : .unwrap(),
339 : )
340 : }
341 :
342 0 : "grpc" => Box::new(
343 0 : GrpcClient::new(
344 0 : &args.page_service_connstring,
345 0 : worker_id.timeline,
346 0 : args.compression,
347 0 : )
348 0 : .await
349 0 : .unwrap(),
350 : ),
351 :
352 0 : scheme => panic!("unsupported scheme {scheme}"),
353 : };
354 0 : run_worker(args, client, ss, cancel, rps_period, ranges, weights).await
355 0 : })
356 0 : };
357 :
358 0 : info!("spawning workers");
359 0 : let mut workers = JoinSet::new();
360 0 : for timeline in timelines.iter().cloned() {
361 0 : for num_client in 0..args.num_clients.get() {
362 0 : let worker_id = WorkerId {
363 0 : timeline,
364 0 : num_client,
365 0 : };
366 0 : workers.spawn(make_worker(worker_id));
367 0 : }
368 : }
369 0 : let workers = async move {
370 0 : while let Some(res) = workers.join_next().await {
371 0 : res.unwrap();
372 0 : }
373 0 : };
374 :
375 0 : info!("waiting for everything to become ready");
376 0 : shared_state.start_work_barrier.wait().await;
377 0 : info!("work started");
378 0 : if let Some(runtime) = args.runtime {
379 0 : tokio::time::sleep(runtime.into()).await;
380 0 : info!("runtime over, signalling cancellation");
381 0 : cancel.cancel();
382 0 : workers.await;
383 0 : info!("work sender exited");
384 : } else {
385 0 : workers.await;
386 0 : unreachable!("work sender never terminates");
387 : }
388 :
389 0 : let output = Output {
390 : total: {
391 0 : let mut agg_stats = request_stats::Stats::new();
392 0 : for stats in all_thread_local_stats.lock().unwrap().iter() {
393 0 : let stats = stats.lock().unwrap();
394 0 : agg_stats.add(&stats);
395 0 : }
396 0 : agg_stats.output()
397 : },
398 : };
399 :
400 0 : let output = serde_json::to_string_pretty(&output).unwrap();
401 0 : println!("{output}");
402 :
403 0 : anyhow::Ok(())
404 0 : }
405 :
406 0 : async fn run_worker(
407 0 : args: &Args,
408 0 : mut client: Box<dyn Client>,
409 0 : shared_state: Arc<SharedState>,
410 0 : cancel: CancellationToken,
411 0 : rps_period: Option<Duration>,
412 0 : ranges: Vec<KeyRange>,
413 0 : weights: rand::distributions::weighted::WeightedIndex<i128>,
414 0 : ) {
415 0 : shared_state.start_work_barrier.wait().await;
416 0 : let client_start = Instant::now();
417 0 : let mut ticks_processed = 0;
418 0 : let mut req_id = 0;
419 0 : let batch_size: usize = args.batch_size.into();
420 :
421 : // Track inflight requests by request ID and start time. This times the request duration, and
422 : // ensures responses match requests. We don't expect responses back in any particular order.
423 : //
424 : // NB: this does not check that all requests received a response, because we don't wait for the
425 : // inflight requests to complete when the duration elapses.
426 0 : let mut inflight: HashMap<u64, Instant> = HashMap::new();
427 :
428 0 : while !cancel.is_cancelled() {
429 : // Detect if a request took longer than the RPS rate
430 0 : if let Some(period) = &rps_period {
431 0 : let periods_passed_until_now =
432 0 : usize::try_from(client_start.elapsed().as_micros() / period.as_micros()).unwrap();
433 :
434 0 : if periods_passed_until_now > ticks_processed {
435 0 : shared_state
436 0 : .live_stats
437 0 : .missed((periods_passed_until_now - ticks_processed) as u64);
438 0 : }
439 0 : ticks_processed = periods_passed_until_now;
440 0 : }
441 :
442 0 : while inflight.len() < args.queue_depth.get() {
443 0 : req_id += 1;
444 0 : let start = Instant::now();
445 0 : let (req_lsn, mod_lsn, rel, blks) = {
446 : /// Converts a compact i128 key to a relation tag and block number.
447 0 : fn key_to_block(key: i128) -> (RelTag, u32) {
448 0 : let key = Key::from_i128(key);
449 0 : assert!(key.is_rel_block_key());
450 0 : key.to_rel_block()
451 0 : .expect("we filter non-rel-block keys out above")
452 0 : }
453 :
454 : // Pick a random page from a random relation.
455 0 : let mut rng = rand::thread_rng();
456 0 : let r = &ranges[weights.sample(&mut rng)];
457 0 : let key: i128 = rng.gen_range(r.start..r.end);
458 0 : let (rel_tag, block_no) = key_to_block(key);
459 :
460 0 : let mut blks = VecDeque::with_capacity(batch_size);
461 0 : blks.push_back(block_no);
462 :
463 : // If requested, populate a batch of sequential pages. This is how Postgres will
464 : // request page batches (e.g. prefetches). If we hit the end of the relation, we
465 : // grow the batch towards the start too.
466 0 : for i in 1..batch_size {
467 0 : let (r, b) = key_to_block(key + i as i128);
468 0 : if r != rel_tag {
469 0 : break; // went outside relation
470 0 : }
471 0 : blks.push_back(b)
472 : }
473 :
474 0 : if blks.len() < batch_size {
475 : // Grow batch backwards if needed.
476 0 : for i in 1..batch_size {
477 0 : let (r, b) = key_to_block(key - i as i128);
478 0 : if r != rel_tag {
479 0 : break; // went outside relation
480 0 : }
481 0 : blks.push_front(b)
482 : }
483 0 : }
484 :
485 : // We assume that the entire batch can fit within the relation.
486 0 : assert_eq!(blks.len(), batch_size, "incomplete batch");
487 :
488 0 : let req_lsn = if rng.gen_bool(args.req_latest_probability) {
489 0 : Lsn::MAX
490 : } else {
491 0 : r.timeline_lsn
492 : };
493 0 : (req_lsn, r.timeline_lsn, rel_tag, blks.into())
494 : };
495 0 : client
496 0 : .send_get_page(req_id, req_lsn, mod_lsn, rel, blks)
497 0 : .await
498 0 : .unwrap();
499 0 : let old = inflight.insert(req_id, start);
500 0 : assert!(old.is_none(), "duplicate request ID {req_id}");
501 : }
502 :
503 0 : let (req_id, pages) = client.recv_get_page().await.unwrap();
504 0 : assert_eq!(pages.len(), batch_size, "unexpected page count");
505 0 : assert!(pages.iter().all(|p| !p.is_empty()), "empty page");
506 0 : let start = inflight
507 0 : .remove(&req_id)
508 0 : .expect("response for unknown request ID");
509 0 : let end = Instant::now();
510 0 : shared_state.live_stats.request_done();
511 0 : ticks_processed += 1;
512 0 : STATS.with(|stats| {
513 0 : stats
514 0 : .borrow()
515 0 : .lock()
516 0 : .unwrap()
517 0 : .observe(end.duration_since(start))
518 0 : .unwrap();
519 0 : });
520 :
521 0 : if let Some(period) = &rps_period {
522 0 : let next_at = client_start
523 0 : + Duration::from_micros(
524 0 : (ticks_processed) as u64 * u64::try_from(period.as_micros()).unwrap(),
525 0 : );
526 0 : tokio::time::sleep_until(next_at.into()).await;
527 0 : }
528 : }
529 0 : }
530 :
531 : /// A benchmark client, to allow switching out the transport protocol.
532 : ///
533 : /// For simplicity, this just uses separate asynchronous send/recv methods. The send method could
534 : /// return a future that resolves when the response is received, but we don't really need it.
535 : #[async_trait]
536 : trait Client: Send {
537 : /// Sends an asynchronous GetPage request to the pageserver.
538 : async fn send_get_page(
539 : &mut self,
540 : req_id: u64,
541 : req_lsn: Lsn,
542 : mod_lsn: Lsn,
543 : rel: RelTag,
544 : blks: Vec<u32>,
545 : ) -> anyhow::Result<()>;
546 :
547 : /// Receives the next GetPage response from the pageserver.
548 : async fn recv_get_page(&mut self) -> anyhow::Result<(u64, Vec<Bytes>)>;
549 : }
550 :
551 : /// A libpq-based Pageserver client.
552 : struct LibpqClient {
553 : inner: pageserver_client::page_service::PagestreamClient,
554 : // Track sent batches, so we know how many responses to expect.
555 : batch_sizes: VecDeque<usize>,
556 : }
557 :
558 : impl LibpqClient {
559 0 : async fn new(connstring: &str, ttid: TenantTimelineId) -> anyhow::Result<Self> {
560 0 : let inner = pageserver_client::page_service::Client::new(connstring.to_string())
561 0 : .await?
562 0 : .pagestream(ttid.tenant_id, ttid.timeline_id)
563 0 : .await?;
564 0 : Ok(Self {
565 0 : inner,
566 0 : batch_sizes: VecDeque::new(),
567 0 : })
568 0 : }
569 : }
570 :
571 : #[async_trait]
572 : impl Client for LibpqClient {
573 0 : async fn send_get_page(
574 : &mut self,
575 : req_id: u64,
576 : req_lsn: Lsn,
577 : mod_lsn: Lsn,
578 : rel: RelTag,
579 : blks: Vec<u32>,
580 0 : ) -> anyhow::Result<()> {
581 : // libpq doesn't support client-side batches, so we send a bunch of individual requests
582 : // instead in the hope that the server will batch them for us. We use the same request ID
583 : // for all, because we'll return a single batch response.
584 0 : self.batch_sizes.push_back(blks.len());
585 0 : for blkno in blks {
586 0 : let req = PagestreamGetPageRequest {
587 0 : hdr: PagestreamRequest {
588 0 : reqid: req_id,
589 0 : request_lsn: req_lsn,
590 0 : not_modified_since: mod_lsn,
591 0 : },
592 0 : rel,
593 0 : blkno,
594 0 : };
595 0 : self.inner.getpage_send(req).await?;
596 : }
597 0 : Ok(())
598 0 : }
599 :
600 0 : async fn recv_get_page(&mut self) -> anyhow::Result<(u64, Vec<Bytes>)> {
601 0 : let batch_size = self.batch_sizes.pop_front().unwrap();
602 0 : let mut batch = Vec::with_capacity(batch_size);
603 0 : let mut req_id = None;
604 0 : for _ in 0..batch_size {
605 0 : let resp = self.inner.getpage_recv().await?;
606 0 : if req_id.is_none() {
607 0 : req_id = Some(resp.req.hdr.reqid);
608 0 : }
609 0 : assert_eq!(req_id, Some(resp.req.hdr.reqid), "request ID mismatch");
610 0 : batch.push(resp.page);
611 : }
612 0 : Ok((req_id.unwrap(), batch))
613 0 : }
614 : }
615 :
616 : /// A gRPC Pageserver client.
617 : struct GrpcClient {
618 : req_tx: tokio::sync::mpsc::Sender<page_api::GetPageRequest>,
619 : resp_rx: Pin<Box<dyn Stream<Item = Result<page_api::GetPageResponse, tonic::Status>> + Send>>,
620 : }
621 :
622 : impl GrpcClient {
623 0 : async fn new(
624 0 : connstring: &str,
625 0 : ttid: TenantTimelineId,
626 0 : compression: bool,
627 0 : ) -> anyhow::Result<Self> {
628 0 : let mut client = page_api::Client::new(
629 0 : connstring.to_string(),
630 0 : ttid.tenant_id,
631 0 : ttid.timeline_id,
632 0 : ShardIndex::unsharded(),
633 0 : None,
634 0 : compression.then_some(tonic::codec::CompressionEncoding::Zstd),
635 0 : )
636 0 : .await?;
637 :
638 : // The channel has a buffer size of 1, since 0 is not allowed. It does not matter, since the
639 : // benchmark will control the queue depth (i.e. in-flight requests) anyway, and requests are
640 : // buffered by Tonic and the OS too.
641 0 : let (req_tx, req_rx) = tokio::sync::mpsc::channel(1);
642 0 : let req_stream = tokio_stream::wrappers::ReceiverStream::new(req_rx);
643 0 : let resp_rx = Box::pin(client.get_pages(req_stream).await?);
644 :
645 0 : Ok(Self { req_tx, resp_rx })
646 0 : }
647 : }
648 :
649 : #[async_trait]
650 : impl Client for GrpcClient {
651 0 : async fn send_get_page(
652 : &mut self,
653 : req_id: u64,
654 : req_lsn: Lsn,
655 : mod_lsn: Lsn,
656 : rel: RelTag,
657 : blks: Vec<u32>,
658 0 : ) -> anyhow::Result<()> {
659 0 : let req = page_api::GetPageRequest {
660 0 : request_id: req_id,
661 0 : request_class: page_api::GetPageClass::Normal,
662 0 : read_lsn: page_api::ReadLsn {
663 0 : request_lsn: req_lsn,
664 0 : not_modified_since_lsn: Some(mod_lsn),
665 0 : },
666 0 : rel,
667 0 : block_numbers: blks,
668 0 : };
669 0 : self.req_tx.send(req).await?;
670 0 : Ok(())
671 0 : }
672 :
673 0 : async fn recv_get_page(&mut self) -> anyhow::Result<(u64, Vec<Bytes>)> {
674 0 : let resp = self.resp_rx.next().await.unwrap().unwrap();
675 0 : anyhow::ensure!(
676 0 : resp.status_code == page_api::GetPageStatusCode::Ok,
677 0 : "unexpected status code: {}",
678 : resp.status_code,
679 : );
680 0 : Ok((resp.request_id, resp.page_images))
681 0 : }
682 : }
|
