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.

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

Optimize the performance by passing MessageID implementations by pointers #968

Merged

Conversation

BewareMyPower
Copy link
Contributor

Motivation

Currently there are three implementations of the MessageID interface:

  • messageID: 24 bytes
  • trackingMessageID: 64 bytes
  • chunkMessageID: 80 bytes

However, for all methods of them, the receiver is a value rather than a pointer. It's inefficient because each time a method is called, the copy would happen.

Reference: https://go.dev/tour/methods/8

Modifications

  • Change the receiver from value to pointer for all MessageID implementations.
  • Use pointers as the returned values and function parameters for these implementations everywhere.

The trackingMessageID.Undefined method is removed because it's never used now. Though it's a public method, the struct and its factory function are not exposed, so I think it's reasonable.

Remove the benchmark added in
#324. The result is obvious and this test is meaningless. I tried passing the trackingMessageID by pointer and the result reduced from 8.548 ns/op to 1.628 ns/op. It's obvious because a pointer is only 8 bytes while a trackingMessageID is 64 bytes. The overhead of accessing by pointers is far less than copying the extra bytes.

@BewareMyPower BewareMyPower self-assigned this Feb 27, 2023
@BewareMyPower BewareMyPower marked this pull request as draft February 27, 2023 16:07
@BewareMyPower BewareMyPower marked this pull request as ready for review February 27, 2023 16:26
@BewareMyPower BewareMyPower changed the title Optimize performance by passing MessageID implementations by pointers [Performance Optimization] Pass MessageID implementations by pointers Feb 27, 2023
@BewareMyPower BewareMyPower changed the title [Performance Optimization] Pass MessageID implementations by pointers Optimize the performance by passing MessageID implementations by pointers Feb 27, 2023
@BewareMyPower BewareMyPower marked this pull request as draft February 28, 2023 02:56
@BewareMyPower BewareMyPower marked this pull request as ready for review February 28, 2023 05:48
@BewareMyPower BewareMyPower force-pushed the bewaremypower/msg-id-pointer branch 2 times, most recently from a83e350 to f7d69df Compare February 28, 2023 08:02
@BewareMyPower
Copy link
Contributor Author

Wait for #970 to be merged first.

### Motivation

Currently there are three implementations of the `MessageID` interface:
- `messageID`: 24 bytes
- `trackingMessageID`: 64 bytes
- `chunkMessageID`: 80 bytes

However, for all methods of them, the receiver is a value rather than a
pointer. It's inefficient because each time a method is called, the copy
would happen.

Reference: https://go.dev/tour/methods/8

### Modifications

- Change the receiver from value to pointer for all `MessageID`
  implementations.
- Use pointers as the returned values and function parameters for these
  implementations everywhere.

The `trackingMessageID.Undefined` method is removed because it's never
used now. Though it's a public method, the struct and its factory
function are not exposed, so I think it's reasonable.

Remove the benchmark added in
apache#324. The result is
obvious and this test is meaningless. I tried passing the
`trackingMessageID` by pointer and the result reduced from 8.548 ns/op
to 1.628 ns/op. It's obvious because a pointer is only 8 bytes while a
`trackingMessageID` is 64 bytes. The overhead of accessing by pointers
is far less than copying the extra bytes.
Copy link
Member

@tisonkun tisonkun left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Generally looks good.

Do you have a bench result on this change? Since you declaim that it's a performance improvement.

Also, why do you remove impl_message_bench_test.go?

@shibd
Copy link
Member

shibd commented Mar 1, 2023

Also, why do you remove impl_message_bench_test.go?

This test has no assertions and should not be in unit tests.

@tisonkun
Copy link
Member

tisonkun commented Mar 1, 2023

This test has no assertions and should not be in unit tests.

I think they're go benches which will be only triggered if you run go test -bench

@BewareMyPower
Copy link
Contributor Author

Do you have a bench result on this change? Since you declaim that it's a performance improvement.

I will paste a result soon. Though it might not be significant.

Also, why do you remove impl_message_bench_test.go?

See my PR description.

Remove the benchmark added in
#324. The result is obvious and this test is meaningless. I tried passing the trackingMessageID by pointer and the result reduced from 8.548 ns/op to 1.628 ns/op. It's obvious because a pointer is only 8 bytes while a trackingMessageID is 64 bytes. The overhead of accessing by pointers is far less than copying the extra bytes.

@BewareMyPower
Copy link
Contributor Author

Test process

I created a branch from master and apply the following patch to disable ACK grouping (so that each MessageID could be passed)

diff --git a/perf/perf-consumer.go b/perf/perf-consumer.go
index 6b6e411..c7648e3 100644
--- a/perf/perf-consumer.go
+++ b/perf/perf-consumer.go
@@ -79,6 +79,7 @@ func consume(consumeArgs *ConsumeArgs, stop <-chan struct{}) {
                Topic:                          consumeArgs.Topic,
                SubscriptionName:               consumeArgs.SubscriptionName,
                EnableBatchIndexAcknowledgment: consumeArgs.EnableBatchIndexAck,
+               AckGroupingOptions:             &pulsar.AckGroupingOptions{MaxSize: 0},
        })

        if err != nil {

Then, build the perf tool.

cd perf
go build

Since this patch mainly affects the MessageID related operations, let's only test the consumer with each message acknowledged.

Start Pulsar 2.11 standalone. Create two subscriptions sub1 and sub2 on the same topic:

./perf consume --enable-batch-index-ack my-topic -s sub1
./perf consume --enable-batch-index-ack my-topic -s sub2

Then, produce some messages and wait for a while and stop the producer.

./perf produce --batching-num-messages=1 --rate=200000 my-topic

Then we can test the catch-up read. First, test without this patch:

go build
./perf consume --enable-batch-index-ack my-topic -s sub1

Then, apply this patch and repeat the steps above:

git merge bewaremypower/msg-id-pointer
go build
# NOTE: sub1 and sub2 have the same initial position
./perf consume --enable-batch-index-ack my-topic -s sub2

Before running the perf tool each time, restart the standalone.

Test result

Without this patch:

INFO[22:39:09.321] Stats - Consume rate: 8350.1 msg/s -   65.2 Mbps
INFO[22:39:19.321] Stats - Consume rate: 9100.0 msg/s -   71.1 Mbps
INFO[22:39:29.321] Stats - Consume rate: 8750.0 msg/s -   68.4 Mbps
INFO[22:39:39.321] Stats - Consume rate: 8870.0 msg/s -   69.3 Mbps
INFO[22:39:49.321] Stats - Consume rate: 9200.0 msg/s -   71.9 Mbps
INFO[22:39:59.321] Stats - Consume rate: 9020.0 msg/s -   70.5 Mbps
INFO[22:40:09.320] Stats - Consume rate: 8659.9 msg/s -   67.7 Mbps
INFO[22:40:19.321] Stats - Consume rate: 8810.1 msg/s -   68.8 Mbps
INFO[22:40:29.321] Stats - Consume rate: 8960.0 msg/s -   70.0 Mbps
INFO[22:40:39.321] Stats - Consume rate: 8550.0 msg/s -   66.8 Mbps
INFO[22:40:49.321] Stats - Consume rate: 8300.0 msg/s -   64.8 Mbps
INFO[22:40:59.322] Stats - Consume rate: 8420.0 msg/s -   65.8 Mbps

With this patch:

INFO[22:42:04.326] Stats - Consume rate: 8340.1 msg/s -   65.2 Mbps
INFO[22:42:14.327] Stats - Consume rate: 9930.0 msg/s -   77.6 Mbps
INFO[22:42:24.327] Stats - Consume rate: 8970.0 msg/s -   70.1 Mbps
INFO[22:42:34.326] Stats - Consume rate: 9410.0 msg/s -   73.5 Mbps
INFO[22:42:44.326] Stats - Consume rate: 9400.0 msg/s -   73.4 Mbps
INFO[22:42:54.326] Stats - Consume rate: 9260.0 msg/s -   72.3 Mbps
INFO[22:43:04.327] Stats - Consume rate: 9362.3 msg/s -   73.1 Mbps
INFO[22:43:14.326] Stats - Consume rate: 9317.7 msg/s -   72.8 Mbps
INFO[22:43:24.326] Stats - Consume rate: 9100.0 msg/s -   71.1 Mbps
INFO[22:43:34.326] Stats - Consume rate: 8900.0 msg/s -   69.5 Mbps
INFO[22:43:44.326] Stats - Consume rate: 8830.0 msg/s -   69.0 Mbps
INFO[22:43:54.327] Stats - Consume rate: 9180.0 msg/s -   71.7 Mbps

After this patch, the throughput is more stable and is a little higher than before.

@tisonkun

Copy link
Member

@tisonkun tisonkun left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for your report.

Although I think the performance is not quite optimized, checking mid, ok looks weird to me.

I'm +1 as it simplifies code without perf regression.

@BewareMyPower BewareMyPower merged commit 42ded0d into apache:master Mar 1, 2023
@BewareMyPower BewareMyPower deleted the bewaremypower/msg-id-pointer branch March 1, 2023 16:04
RobertIndie pushed a commit that referenced this pull request Mar 7, 2023
### Motivation

`toTrackingMessageID()` is a function that is widely used in `consumer` implementation. It can convert an interface type  `MessageID` into an struct type `trackingMessageID`. In addition, the second return value also plays a role in checking the `MessageID` type. In other words, it indicates that `MessageID` **cannot** be a user-defined type. From the perspective of code readability, `toTrackingMessageID()` should not do both.

**Note**: After #968 , `toTrackingMessageID()` returns only a pointer now. The role of original `ok` is replaced by nil pointer  now. However, the main content discussed in this PR has not changed.

For example.

https://github.com/apache/pulsar-client-go/blob/e2ea255052e8a527091791ef368851d885ee2d45/pulsar/consumer_regex.go#L176-L181

This example is the correct usage. The `ok` returned by `toTrackingMessageID()` is used to reject user-defined `MessageID`.

https://github.com/apache/pulsar-client-go/blob/e2ea255052e8a527091791ef368851d885ee2d45/pulsar/consumer_partition.go#L470-L473

This example is a bit vague. The actual effect here is the same as the previous example. But it return an error `failed to convert trackingMessageID` which is confusing. 

https://github.com/apache/pulsar-client-go/blob/e2ea255052e8a527091791ef368851d885ee2d45/pulsar/consumer_partition.go#L1816-L1820

In this case. We just want to convert `MessageID` into `trackingMessageID`. We do not care what it really is because it's not possible an invalid `MessageID` implementation.

So, original `toTrackingMessageID()` needs to require a careful look to use it correctly. I think it would be better to split it into two different method. `toTrackingMessageID()` just do the struct conversion, which it's more clearly. And when the new messageID type is added, we can just modify the `checkMessageIDType`.

### Modifications

- Refactor the `toTrackingMessageID()`
- Add the `checkMessageIDType()` to check whether `MessageID` is user-defined.
@RobertIndie RobertIndie added this to the v0.10.0 milestone Mar 27, 2023
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

5 participants