Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[dst] Improve tail-latency for operations of transactions using wait queues #13580

Closed
robertsami opened this issue Aug 11, 2022 · 2 comments
Closed

[dst] Improve tail-latency for operations of transactions using wait queues #13580

robertsami opened this issue Aug 11, 2022 · 2 comments
Assignees
@robertsami
Labels
area/docdb YugabyteDB core features kind/enhancement This is an enhancement of an existing feature priority/medium Medium priority issue

Comments

@robertsami
Copy link
Contributor

robertsami commented Aug 11, 2022 •
edited
Loading

Jira Link: DB-3158

The biggest contributor to higher tail latency is caused by the following case of starvation -- in case there is a high-degree of contention, waiting transactions may get starved by incoming operations which contend for the same latch. We currently have no mechanism to prevent this, which can lead to high tail-latency in some workloads.

Less critically, our process for determining which waiters can be resumed and subsequently resuming them could be improved in a couple ways:

  1. We currently iterate over each of the blocker's waiters and separately acquire a write lock on a mutex to remove the waiter from waiter_status_ before resuming the waiter. We need not re-acquire this write lock for every waiter and can simply acquire it once
  2. We currently resume waiters in the order they arrived, and in serial on a single thread. It might be better to understand which of the waiters will conflict with each other, and then either:
    a. Resolve the first-in waiter and all non-conflicting other waiters in parallel
    b. Resolve the largest set of non-conflicting waiters in parallel, then the second largest, etc
@robertsami robertsami self-assigned this Aug 11, 2022
@robertsami robertsami added the area/docdb YugabyteDB core features label Aug 11, 2022
@yugabyte-ci yugabyte-ci added kind/bug This issue is a bug priority/medium Medium priority issue labels Aug 11, 2022
This was referenced Aug 11, 2022
Closed
Closed
@yugabyte-ci yugabyte-ci added kind/enhancement This is an enhancement of an existing feature and removed kind/bug This issue is a bug labels Oct 19, 2022
@robertsami robertsami changed the title [dst] Improve the continuation of waiters in wait_queue.cc [dst] Improve tail-latency for operations of transactions using wait queues Jan 17, 2023
robertsami added a commit to robertsami/yugabyte-db that referenced this issue Feb 28, 2023
@robertsami
WIP [yugabyte#13580] Implement fairness in wait queues
3baad30 
Summary: tbd

Test Plan: Jenkins

Subscribers: bogdan

Differential Revision: https://phabricator.dev.yugabyte.com/D22968
robertsami added a commit that referenced this issue Mar 1, 2023
@robertsami
[#13580] DocDB: Improve fairness in wait queues
f69dc2a 
Summary:
The main contribution of this revision is to drastically improve p99 performance of workloads using wait-on-conflict concurrency control under high contention, without harming p50 or average performance under normal amounts of contention. We achieve this by making the following improvements:
1. Force incoming requests to check the wait queue once for active blockers, to ensure incoming requests cannot starve waiting transactions which are racing to exit the wait queue
2. Assign serial numbers to incoming requests, and whenever a batch of waiters can be resumed at the same time, ensure they are resumed roughly in the order in which they arrived at the tserver

Additional enhancements include:
1. Reduce copying by consolidating on using TransactionData everywhere, which is pulled into a conflict_data.h file with associated data structures
2. Populate granular intent information on a sub-transaction basis for use by the wait queue
3. Piggy-back off transaction status request in conflict resolution to obtain status tablet info

Test Plan:
Performance was tested on a 16 core 32gb ram alma8 server with a full-LTO release build. In both cases we used the following set-up:
```
create table test (k int primary key, v int);
insert into test select generate_series(0, 11), 0;
```
In both cases, we also ran ysql_bench as follows:
```
build/latest/postgres/bin/ysql_bench --transactions=2000 --jobs=16 --client=16 --file=workload.sql --progress=1 --debug=fails
```

= First test: Max contention =
`workload.sql`
```
begin;
select * from test where k=1 for update;
commit;
```

Baseline:
```
latency average = 19.779 ms
latency stddev = 26.684 ms
tps = 792.780284 (including connections establishing)
tps = 793.793930 (excluding connections establishing)
```

With revision:
```
latency average = 22.632 ms
latency stddev = 3.266 ms
tps = 705.108285 (including connections establishing)
tps = 705.914647 (excluding connections establishing)
```

= Second test: Normal contention =
`workload.sql`
```
begin;
with rand as (select floor(random() * 10 + 1)::int as k) select * from test join rand on rand.k=test.k for update;
commit;
```

Baseline:
```
latency average = 7.317 ms
latency stddev = 6.516 ms
tps = 2117.437801 (including connections establishing)
tps = 2126.594897 (excluding connections establishing)
```

With revision:
```
latency average = 7.055 ms
latency stddev = 5.124 ms
tps = 2236.062486 (including connections establishing)
tps = 2244.260708 (excluding connections establishing)
```

==Takeaways==
1. stddev of latency is substantially improved with this revision, at the expense of a 25% drop in throughput and 35% increase in latency
2. Throughput is not significantly changed in normal contention case

Reviewers: pjain, bkolagani, sergei

Reviewed By: sergei

Subscribers: mbautin, rthallam, bogdan

Differential Revision: https://phabricator.dev.yugabyte.com/D22968
@rthallamko3
Copy link
Contributor

rthallamko3 commented Mar 7, 2023 •
edited
Loading

@robertsami , Can we close this or were you planning additional followups?

@robertsami
Copy link
Contributor Author

robertsami commented Mar 10, 2023 •
edited
Loading

the main fix for fairness has landed in f69dc2a

regarding the remaining minor points in the description:

  1. is no longer relevant
  2. is captured here and not as critical: [DocDB] Avoid re-running conflict resolution for waiters which are still blocked #16389

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@robertsami robertsami

Labels
area/docdb YugabyteDB core features kind/enhancement This is an enhancement of an existing feature priority/medium Medium priority issue
Projects
Wait-Queue Based Locking
Status: Done
Milestone
No milestone
Development

No branches or pull requests
3 participants
@robertsami @rthallamko3 @yugabyte-ci