JetStreams poor performance even with partitioning #5551
Replies: 8 comments 20 replies
-
|
tagging @mtmk who can take a look at the nats.net.v2 publisher code |
Beta Was this translation helpful? Give feedback.
-
|
re .NET client errors, we did have a fix around subscription / inbox memory management in 2.2.3 could you try that please? |
Beta Was this translation helpful? Give feedback.
-
Http client, for instance throws such exception when timeout occurs. I expected something like this here, because the apps code doesn't fire any cancellation tokens. I'll give 2.2.3 a shot, thanks! |
Beta Was this translation helpful? Give feedback.
-
|
I've just checked 2.2.3. Unfortunately it produces the same errors. |
Beta Was this translation helpful? Give feedback.
-
|
thank you. ok it might be a bug or we might be hitting timeouts somehow as you suggested. it would also be nice to see how the tasks are stacking up. we might need to look into creating the new parallel jetstream publish api sooner than later. do you mind creating an issue in https://github.com/nats-io/nats.net.v2/issues with repro code and instructions please? that would be really helpful, unfortunately now I can't see how it's getting into that state from those log entries. thank you in advance 💯 |
Beta Was this translation helpful? Give feedback.
-
|
Repro code and instructions are here. Repro settings for the app: I'll create a separete issue a bit later. |
Beta Was this translation helpful? Give feedback.
-
|
Do you use async publishing from the clients? |
Beta Was this translation helpful? Give feedback.
-
|
Do you use async publishing from the clients? |
Beta Was this translation helpful? Give feedback.
-
In my case i need per subject strict ordering, at the same time i have a lot of different subjects (hundreds of thousands). So, for instance 10 pods have 1280 parallel non-io-blocking publishers. |
Beta Was this translation helpful? Give feedback.
-
|
These numbers seem low, I can do a simple Are the streams replicated? |
Beta Was this translation helpful? Give feedback.
-
|
Yes. R3 replicated. (As it supposed to be in production). |
Beta Was this translation helpful? Give feedback.
-
|
ok if that is the case, then the main dominant issue, since you use sync publishes, will be how long the fastest follower responds to the appendEntry from the NRG (NATS raft group). When you do async we can batch multiple ones into AppendEntries. Some batching can also occur with multiple sync publishers on different connections. So what that means is if the RTT say between two server is 10ms, then the theoretical maximum throughput for sync publishes (meaning one message per AppendEntry) will be 1000/s. In this case partitions are probably not doing anything or very little. |
Beta Was this translation helpful? Give feedback.
-
|
ok if that is the case, then the main dominant issue, since you use sync
publishes, will be how long the fastest follower responds to the
appendEntry from the NRG (NATS raft group).
When you do async we can batch multiple ones into AppendEntries. Some
batching can also occur with multiple sync publishers on different
connections.
So what that means is if the RTT say between two server is 10ms, then the
theoretical maximum throughput for sync publishes (meaning one message per
AppendEntry) will be 1000/s.
In this case partitions are probably not doing anything or very little.
…On Mon, Jun 17, 2024 at 12:33 PM Anton Smolkov ***@***.***> wrote:
@mtmk <https://github.com/mtmk> Thanks for your response.
As i answered below i use synchronous publishing, but from 128 tasks in
parallel.
So it’s equivalent to batch size of 128 from your example.
I tried to expand the amount of in-flight parallel tasks, but with no
success.
I’ll check it again tomorrow and publish results here, but as you can see
- even with current parallelism level NATS brokers CPUs are fully utilized.
—
Reply to this email directly, view it on GitHub
<#5551 (reply in thread)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAAV74NVAGTJLC3WBS4BWLDZH42ZHAVCNFSM6AAAAABJN6JDICVHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4TOOJYGM3DA>
.
You are receiving this because you commented.Message ID:
***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
Does that mean that async and sync publish are different on the protocol level? As far as i understand in dotnet client for now we have only one kind of publish wrapped to the task(future) Btw, I do use dedicated connection per each publishing task (so 1280 connections from the example above).
Is it still a case even if there is 128 separate r3 JetStreams distributed on 14 nodes? |
Beta Was this translation helpful? Give feedback.
-
|
The NATS server can batch multiple messages to the NRG layer when pulling from the internal proposal queue. So you can see some benefit there in msgs/sec. Point is that the limiting factor will be in the RTT between leader and fastest follower. If the inbound rate is more important than you could go to async replication with the ingest stream being an R1 and having a mirror, etc. Of course that does not help with server restarts etc. You could also look for clusters that have lower rtt between servers. |
Beta Was this translation helpful? Give feedback.
-
Could you please ellaborate a little bit more: |
Beta Was this translation helpful? Give feedback.
-
|
There is something not right about your setup because those numbers are not very high for such a large deployment and it’s also strange that you manage to max out a total of 70 CPU cores on all your servers.
I would do some tests varying parameters to try and investigate: eg try for the streams using memory storage, try with replicas=1, try using “nats bench” and see if you see similar numbers or not.
|
Beta Was this translation helpful? Give feedback.
-
|
I already use im-memory storage for JetStreams. |
Beta Was this translation helpful? Give feedback.
-
|
Next step would be to schedule a call so we can triage your system. |
Beta Was this translation helpful? Give feedback.
-
|
We can schedule the call if necessary. But there is 10 hours timezone difference, so it's not going to be easy to find appropriate time |
Beta Was this translation helpful? Give feedback.
-
|
Also be good try against RC5 of next release. https://github.com/nats-io/nats-server/releases/tag/v2.10.17-RC.5 |
Beta Was this translation helpful? Give feedback.
-
Didn't find corresponding docker images. I see only nightly docker builds from |
Beta Was this translation helpful? Give feedback.
-
|
|
Beta Was this translation helpful? Give feedback.
-
|
nats bench: sync: ~ 3500, 70 cores The results are similar to 2.10.16. A bit better for async |
Beta Was this translation helpful? Give feedback.
-
Hi there.
During my performance evaluations of JetStream, I noticed subpar results in terms of publishing efficiency and scalability.
It appears that overall publishing performace is low and JetStream doesn't achieve linear scalability through partitioning.
Could this be standard behavior, and might you suggest strategies to enhance JetStream's performance?
Publishing code (Just in case)
https://gist.github.com/AntonSmolkov/e561fa2bf92eaab9c1423c2f818754ed
Envinronment:
v1.25.4(public cloud, Intel® Xeon® Gold 6230 (Cascade Lake))2.10.16-alpineNATS.Client.JetStream 2.2.2Settings:
GOMEMLIMIT: 4000MiB100 bytes.JetStreams settings details
Results TL;DR
128 R3 JetStreams overall performace:
Core NATS Pub/Sub performance:
Round 1
Specs
Graphs
expand
NATS:
Producer:
RPS:
Result
Overall result across all broker and producer instances is about 26000 RPS
Broker instaces use ~100% CPU and ~70% RAM resources (25 cores, 18 GiB).
Producer instances use ~70% of CPU and ~30% of RAM (7 CPU cores, 3 GiB)
nats stream reportsconfirms that messages are distributed evenlyexpand
Obtaining Stream statsRound 2
Specs
Graphs
expand
NATS:
Producer:
RPS:
Result
Overall result across all broker and producer instances is about 38000 RPS
Broker instaces ~100% CPU and ~70% RAM (50 cores, 33 GiB).
Producer instances use ~60% of CPU and ~25% of RAM (12 CPU cores, 4.6 GiB)
nats stream reportsconfirms that messages are distributed evenlyexpand
Round 3
Specs
Graphs
expand
NATS:
Producer:
RPS:
Result
Overall result across all broker and producer instances is about 43500 RPS
Broker instaces use ~100% CPU and ~70% RAM (70 cores, 46 GiB).
Producer instances use ~60% of CPU and ~25% of RAM (15 CPU cores, 6 GiB)
nats stream reportsconfirms that messages are distributed evenlyexpand
Round 4 (core pub/sub, for comparison)
(All JetStreams were deleted beforehand)
Specs
Graphs
expand
NATS:
Producer:
RPS:
Result
Overall result across all broker and producer instances is about 7_620_000 (7 million!) RPS
Broker instaces use ~20% CPU and ~6% RAM resources (5 cores, 1.4 GiB).
Producer instances use ~100% of CPU and ~10% of RAM (20 CPU cores, 2 GiB)
Update: NATS Bench resulsts
128 JetStream partitions were set up using this guide
Specs
Nats bench 1 (sync)
Graphs
expand
Result
Overall result across all brokers instances is about 34560 RPS
Broker instaces use ~100% CPU and ~40% RAM (70 cores, 29 GiB).
Producer instance(
nats-box) uses ~60% of CPU and ~1% of RAM (3 CPU cores, 241 MiB)nats bench 2 (async)
Graphs
expand
Result
Overall result across all brokers instances is about 318720 RPS
Broker instaces use ~90% CPU and ~65% RAM (68 cores, 49 GiB).
Producer instance(
nats-box) uses ~80% of CPU and ~25% of RAM (4 CPU cores, 1.35 GiB)Beta Was this translation helpful? Give feedback.
All reactions