Use sync.Pool to share per-connection transport write buffer. #6309

s-matyukevich · 2023-05-23T02:03:31Z

This is an attempt to fix #5759.

The use-case where this fix could be especially useful is a server with a lot of mostly idle connections (like connections that handle only healthchecks) In our infrastructure this use-case is very common, as we have some servers that handle request from thousands of clients, all of which generate relatively small volume of requests but has an active healthcheck setup, which forces every client connect to every server. Some kind of subsetting would be useful in such scenario, but gRPC doesn't really implement subsetting. Grpc-go isn't great at handling memory in such use-case. Right now a lot of our users put envoy in front of grpc server just because it handles memory per connection a lot more efficiently (in one case we were able to reduce the memory consumed by an app from 1.5 GB to 500 MB just by putting envoy in front of the grpc server)

This PR adds sync.Pool to the transport layer to share the write buffer between connections (I plan to provide a similar fix for the read buffer if this PR will be accepted) The buffer is claimed in the Write method and released on Flush, with an additional optimization that preserves the buffer if a single Write results in multiple Flushes.

Testing

I submitted #6299 so that I can reproduce the case described above with benchmarks. I used the following comnand to demonstrate the impact of this change

go run benchmark/benchmain/main.go -workloads all -kbps 0 -latency 0s -maxConcurrentCalls 1000 -reqSizeBytes 1 -respSizeBytes 1 -benchtime 1m -shareConnection=false -sleepBetweenRPCs=20s -resultFile /tmp/master1 -memProfile /tmp/master1-mem

As you can see, I used tiny requests with a combination of large delays between requests to minimize the impact of other allocations on memory stats.

Here are the results that I got:

Without the optimization:

root@traffic-grpc-go-benchmark-manual:/tmp/grpc-go# go tool pprof -top -inuse_space ../master3-mem
File: main
Type: inuse_space
Time: May 23, 2023 at 12:36am (UTC)
Showing nodes accounting for 104.04MB, 95.85% of 108.55MB total
Dropped 24 nodes (cum <= 0.54MB)
      flat  flat%   sum%        cum   cum%
   59.04MB 54.39% 54.39%    59.04MB 54.39%  google.golang.org/grpc/internal/transport.newBufWriter (inline)
   32.49MB 29.94% 84.33%    32.49MB 29.94%  bufio.NewReaderSize (inline)
    5.50MB  5.07% 89.40%     5.50MB  5.07%  runtime.malg
       2MB  1.84% 91.24%        2MB  1.84%  google.golang.org/grpc/internal/transport.(*http2Server).operateHeaders
    1.50MB  1.38% 92.62%     1.50MB  1.38%  sync.(*Pool).pinSlow
       1MB  0.92% 93.55%        1MB  0.92%  bytes.growSlice
       1MB  0.92% 94.47%        1MB  0.92%  golang.org/x/net/http2/hpack.(*headerFieldTable).addEntry (inline)
    0.50MB  0.46% 94.93%    94.04MB 86.63%  google.golang.org/grpc/internal/transport.NewServerTransport
    0.50MB  0.46% 95.39%        1MB  0.92%  golang.org/x/net/http2/hpack.NewDecoder
    0.50MB  0.46% 95.85%     1.50MB  1.38%  bytes.(*Buffer).grow
         0     0% 95.85%        1MB  0.92%  bytes.(*Buffer).ReadFrom
         0     0% 95.85%        1MB  0.92%  golang.org/x/net/http2.(*Framer).ReadFrame
         0     0% 95.85%        1MB  0.92%  golang.org/x/net/http2.(*Framer).readMetaFrame
         0     0% 95.85%        1MB  0.92%  golang.org/x/net/http2/hpack.(*Decoder).Write
         0     0% 95.85%        1MB  0.92%  golang.org/x/net/http2/hpack.(*Decoder).parseFieldLiteral
         0     0% 95.85%        1MB  0.92%  golang.org/x/net/http2/hpack.(*Decoder).parseHeaderFieldRepr
         0     0% 95.85%        1MB  0.92%  golang.org/x/net/http2/hpack.(*dynamicTable).add
         0     0% 95.85%        1MB  0.92%  google.golang.org/grpc.(*Server).Serve
         0     0% 95.85%    94.04MB 86.63%  google.golang.org/grpc.(*Server).Serve.func3
         0     0% 95.85%    94.04MB 86.63%  google.golang.org/grpc.(*Server).handleRawConn
         0     0% 95.85%        5MB  4.61%  google.golang.org/grpc.(*Server).handleRawConn.func1
         0     0% 95.85%        1MB  0.92%  google.golang.org/grpc.(*Server).handleStream
         0     0% 95.85%    94.04MB 86.63%  google.golang.org/grpc.(*Server).newHTTP2Transport
         0     0% 95.85%        1MB  0.92%  google.golang.org/grpc.(*Server).processStreamingRPC
         0     0% 95.85%        5MB  4.61%  google.golang.org/grpc.(*Server).serveStreams
         0     0% 95.85%        1MB  0.92%  google.golang.org/grpc.(*Server).serveStreams.func1.1
         0     0% 95.85%        1MB  0.92%  google.golang.org/grpc/benchmark/latency.(*conn).Read
         0     0% 95.85%        1MB  0.92%  google.golang.org/grpc/benchmark/latency.(*listener).Accept
         0     0% 95.85%     1.50MB  1.38%  google.golang.org/grpc/internal/transport.(*bufferPool).get (inline)
         0     0% 95.85%        5MB  4.61%  google.golang.org/grpc/internal/transport.(*http2Server).HandleStreams
         0     0% 95.85%     1.50MB  1.38%  google.golang.org/grpc/internal/transport.(*http2Server).handleData
         0     0% 95.85%        1MB  0.92%  google.golang.org/grpc/internal/transport.NewServerTransport.func2
         0     0% 95.85%    92.54MB 85.25%  google.golang.org/grpc/internal/transport.newFramer
         0     0% 95.85%        1MB  0.92%  io.Copy (inline)
         0     0% 95.85%        1MB  0.92%  io.CopyN
         0     0% 95.85%        1MB  0.92%  io.ReadAtLeast
         0     0% 95.85%        1MB  0.92%  io.ReadFull (inline)
         0     0% 95.85%        1MB  0.92%  io.copyBuffer
         0     0% 95.85%        1MB  0.92%  net.(*TCPListener).Accept
         0     0% 95.85%        1MB  0.92%  net.(*TCPListener).accept
         0     0% 95.85%     5.50MB  5.07%  runtime.newproc.func1
         0     0% 95.85%     5.50MB  5.07%  runtime.newproc1
         0     0% 95.85%     5.50MB  5.07%  runtime.systemstack
         0     0% 95.85%     1.50MB  1.38%  sync.(*Pool).Get
         0     0% 95.85%     1.50MB  1.38%  sync.(*Pool).pin

With the optimization:

root@traffic-grpc-go-benchmark-manual:/tmp/grpc-go# go tool pprof -top -inuse_space ../fork3-mem
File: main
Type: inuse_space
Time: May 23, 2023 at 12:42am (UTC)
Showing nodes accounting for 55.53MB, 100% of 55.53MB total
      flat  flat%   sum%        cum   cum%
   31.98MB 57.59% 57.59%    31.98MB 57.59%  bufio.NewReaderSize (inline)
    8.50MB 15.31% 72.90%     8.50MB 15.31%  runtime.malg
    2.50MB  4.51% 77.41%     2.50MB  4.51%  sync.(*Pool).pinSlow
       2MB  3.60% 81.01%        2MB  3.60%  golang.org/x/net/http2/hpack.(*headerFieldTable).addEntry (inline)
    1.50MB  2.70% 83.71%        3MB  5.40%  google.golang.org/grpc/internal/transport.(*http2Server).operateHeaders
       1MB  1.80% 85.52%        1MB  1.80%  bytes.growSlice
       1MB  1.80% 87.32%        1MB  1.80%  context.(*cancelCtx).Done
       1MB  1.80% 89.12%    35.48MB 63.89%  google.golang.org/grpc/internal/transport.NewServerTransport
       1MB  1.80% 90.92%        1MB  1.80%  google.golang.org/grpc/internal/transport.newControlBuffer (inline)
    0.53MB  0.96% 91.88%     0.53MB  0.96%  google.golang.org/grpc/internal/transport.newFramer.func1
    0.50MB  0.91% 92.79%     0.50MB  0.91%  runtime.doaddtimer
    0.50MB  0.91% 93.70%     0.50MB  0.91%  google.golang.org/protobuf/reflect/protoregistry.(*Files).RegisterFile.func2
    0.50MB   0.9% 94.60%     0.50MB   0.9%  google.golang.org/grpc.(*Server).processStreamingRPC
    0.50MB   0.9% 95.50%     0.50MB   0.9%  golang.org/x/net/http2.NewFramer
    0.50MB   0.9% 96.40%     0.50MB   0.9%  golang.org/x/net/http2.NewFramer.func2
    0.50MB   0.9% 97.30%     0.50MB   0.9%  net.newFD (inline)
    0.50MB   0.9% 98.20%     0.50MB   0.9%  google.golang.org/grpc/internal/transport.(*loopyWriter).registerStreamHandler (inline)
    0.50MB   0.9% 99.10%     0.50MB   0.9%  google.golang.org/grpc/internal/transport.newLoopyWriter
    0.50MB   0.9%   100%     0.50MB   0.9%  context.WithValue
         0     0%   100%        1MB  1.80%  bytes.(*Buffer).ReadFrom
         0     0%   100%        1MB  1.80%  bytes.(*Buffer).grow
         0     0%   100%     1.50MB  2.70%  golang.org/x/net/http2.(*Framer).ReadFrame
         0     0%   100%     0.53MB  0.96%  golang.org/x/net/http2.(*Framer).WriteHeaders
         0     0%   100%     0.53MB  0.96%  golang.org/x/net/http2.(*Framer).endWrite
         0     0%   100%        1MB  1.80%  golang.org/x/net/http2.(*Framer).readMetaFrame
         0     0%   100%        1MB  1.80%  golang.org/x/net/http2/hpack.(*Decoder).Write
         0     0%   100%        1MB  1.80%  golang.org/x/net/http2/hpack.(*Decoder).parseFieldLiteral
         0     0%   100%        1MB  1.80%  golang.org/x/net/http2/hpack.(*Decoder).parseHeaderFieldRepr
         0     0%   100%        1MB  1.80%  golang.org/x/net/http2/hpack.(*Encoder).WriteField
         0     0%   100%        2MB  3.60%  golang.org/x/net/http2/hpack.(*dynamicTable).add
         0     0%   100%     0.50MB   0.9%  google.golang.org/grpc.(*Server).Serve
         0     0%   100%    35.48MB 63.89%  google.golang.org/grpc.(*Server).Serve.func3
         0     0%   100%    35.48MB 63.89%  google.golang.org/grpc.(*Server).handleRawConn
         0     0%   100%        6MB 10.81%  google.golang.org/grpc.(*Server).handleRawConn.func1
         0     0%   100%     0.50MB   0.9%  google.golang.org/grpc.(*Server).handleStream
         0     0%   100%    35.48MB 63.89%  google.golang.org/grpc.(*Server).newHTTP2Transport
         0     0%   100%        6MB 10.81%  google.golang.org/grpc.(*Server).serveStreams
         0     0%   100%     0.50MB   0.9%  google.golang.org/grpc.(*Server).serveStreams.func1.1
         0     0%   100%        1MB  1.80%  google.golang.org/grpc/benchmark/latency.(*conn).Read
         0     0%   100%     0.50MB   0.9%  google.golang.org/grpc/benchmark/latency.(*listener).Accept
         0     0%   100%     0.53MB  0.96%  google.golang.org/grpc/internal/transport.(*bufWriter).Write
         0     0%   100%     2.50MB  4.51%  google.golang.org/grpc/internal/transport.(*bufferPool).get (inline)
         0     0%   100%        1MB  1.80%  google.golang.org/grpc/internal/transport.(*http2Client).handleData
         0     0%   100%        1MB  1.80%  google.golang.org/grpc/internal/transport.(*http2Client).reader
         0     0%   100%        6MB 10.81%  google.golang.org/grpc/internal/transport.(*http2Server).HandleStreams
         0     0%   100%     1.50MB  2.70%  google.golang.org/grpc/internal/transport.(*http2Server).handleData
         0     0%   100%     2.03MB  3.66%  google.golang.org/grpc/internal/transport.(*loopyWriter).handle
         0     0%   100%     1.53MB  2.76%  google.golang.org/grpc/internal/transport.(*loopyWriter).headerHandler
         0     0%   100%     0.53MB  0.96%  google.golang.org/grpc/internal/transport.(*loopyWriter).originateStream
         0     0%   100%     2.03MB  3.66%  google.golang.org/grpc/internal/transport.(*loopyWriter).run
         0     0%   100%     1.53MB  2.76%  google.golang.org/grpc/internal/transport.(*loopyWriter).writeHeader
         0     0%   100%        2MB  3.60%  google.golang.org/grpc/internal/transport.NewServerTransport.func2
         0     0%   100%    32.48MB 58.49%  google.golang.org/grpc/internal/transport.newFramer
         0     0%   100%     0.53MB  0.96%  google.golang.org/grpc/internal/transport.newHTTP2Client.func6
         0     0%   100%     0.50MB  0.91%  google.golang.org/grpc/interop/grpc_testing.file_grpc_testing_control_proto_init
         0     0%   100%     0.50MB  0.91%  google.golang.org/grpc/interop/grpc_testing.init.1
         0     0%   100%     0.50MB   0.9%  google.golang.org/grpc/metadata.NewIncomingContext (inline)
         0     0%   100%     0.50MB  0.91%  google.golang.org/protobuf/internal/filedesc.Builder.Build
         0     0%   100%     0.50MB  0.91%  google.golang.org/protobuf/internal/filetype.Builder.Build
         0     0%   100%     0.50MB  0.91%  google.golang.org/protobuf/reflect/protoregistry.(*Files).RegisterFile
         0     0%   100%     0.50MB  0.91%  google.golang.org/protobuf/reflect/protoregistry.rangeTopLevelDescriptors
         0     0%   100%        1MB  1.80%  io.Copy (inline)
         0     0%   100%        1MB  1.80%  io.CopyN
         0     0%   100%        1MB  1.80%  io.ReadAtLeast
         0     0%   100%        1MB  1.80%  io.ReadFull (inline)
         0     0%   100%        1MB  1.80%  io.copyBuffer
         0     0%   100%     0.50MB   0.9%  net.(*TCPListener).Accept
         0     0%   100%     0.50MB   0.9%  net.(*TCPListener).accept
         0     0%   100%     0.50MB   0.9%  net.(*netFD).accept
         0     0%   100%     0.50MB  0.91%  runtime.doInit
         0     0%   100%     0.50MB  0.91%  runtime.main
         0     0%   100%     0.50MB  0.91%  runtime.mcall
         0     0%   100%     0.50MB  0.91%  runtime.modtimer
         0     0%   100%     8.50MB 15.31%  runtime.newproc.func1
         0     0%   100%     8.50MB 15.31%  runtime.newproc1
         0     0%   100%     0.50MB  0.91%  runtime.park_m
         0     0%   100%     0.50MB  0.91%  runtime.resetForSleep
         0     0%   100%     0.50MB  0.91%  runtime.resettimer (inline)
         0     0%   100%     8.50MB 15.31%  runtime.systemstack
         0     0%   100%     3.03MB  5.46%  sync.(*Pool).Get
         0     0%   100%     2.50MB  4.51%  sync.(*Pool).pin

As you can see, in the unoptimized benchmark run newBufWriter method allocates 59.04MB or 54.39% of the whole memory. In the optimized run we now see this line instead 0.53MB 0.96% 91.88% 0.53MB 0.96% google.golang.org/grpc/internal/transport.newFramer.func1 which corresponds to the memory allocated by the sync.Pool that I added to the newFramer function.

I also have some results that demonstrates that this change doesn't negatively impact performance in the more realistic benchmark setups, but I will provide those results a bit later.

RELEASE NOTES:

client & server: Add experimental [With]SharedWriteBuffer to improve performance by reducing allocations when sending RPC messages. (Disabled by default.)

s-matyukevich · 2023-05-23T02:11:03Z

internal/transport/http_util.go

+		}
+		writeBufferPoolMap[size] = pool
+	}
+	writeBufferMutex.Unlock()


In theory, we can get rid of this mutex and pool map if we store sync.Pool per ClientConnection and per server. However, this will require modifying a lot of internal interface to pass the pool all the way down to the transport layer. This also will prevent us from sharing the pool between different ClientConnections and servers, so I decided to take the mutex approach.

As far as I understand, this mutex shouldn't really impact the performance, as it got called only on new connection creation, which happens not very often in http2.

When do entries from this map ever get deleted? Can this map grow in a unbounded fashion?

Right now - never. In theory it can grow unbounded if the app creates multiple servers and/or ClientConnections with different buffer sizes. I can't think about any use-case where this could be useful, but it is possible. I am not sure if it worth adding more complicated logic to protect against this, because my understanding is that go runtime will eventually deallocate all memory from unused pools, so we will be waisting memory only on storing pointers to empty pools, which should be neglectable in practise.

Still I can fix this if you prefer:

I can add a goroutine that removes unused pools. I'll probably need to store something like lastModifedTime for each pool to do that. The downsides are some additional complexity + the fact that we need to take the lock in the goroutine.

I can store a pool per Server and ClientConnection and propagate it all the way down to the transport layer. The downsides are the fact that we need to extend some internal interfaces + we won't be able to reuse pools between different connections that uses the same buffer size (which I assume is a very common use-case) But in this case we can get rid of the lock completely.

Do you have any other ideas? Let me know what I should do here.

In theory it can grow unbounded if the app creates multiple servers and/or ClientConnections with different buffer sizes. I can't think about any use-case where this could be useful, but it is possible.

I agree. I don't expect applications to create multiple grpc.ClientConns or grpc.Servers and pass them different values for the send and receive buffer sizes. But given that we are going to have a dial option and a server option to enable the sync.Pool on the client and server respectively, we could document it there, just so no one complains saying a few bytes are leaked.

@dfawley : Thoughts/Objections?

internal/transport/http_util.go

s-matyukevich · 2023-05-26T21:29:39Z

I was trying to test whether this change causes any significant CPU/latency regressions and run benchmarks on a dedicated 8 CPU 32GB virtual machine. I used #6299 with the following parameters

go run benchmark/benchmain/main.go -workloads all -kbps 0 -latency 0s -maxConcurrentCalls 50 -reqSizeBytes 1024,102400 -respSizeBytes 1024,102400 -benchtime 1m -connections=50 -sleepBetweenRPCs=0s,10s -resultFile /tmp/master1 -memProfile /tmp/master1-mem -memProfileRate 1024

Here are the results.

s-matyukevich · 2023-05-30T21:27:41Z

I tested this with a real app and got 40% memory reduction. Each instance of the app is handling ~1.2k connections and has low volume of requests.

dfawley · 2023-05-31T20:15:22Z

Unfortunately with @arvindbr8 and @easwars out of town we are a bit short staffed right now. Will resume in ~2 weeks if that's okay, thanks!

s-matyukevich · 2023-05-31T20:27:42Z

Unfortunately with @arvindbr8 and @easwars out of town we are a bit short staffed right now. Will resume in ~2 weeks if that's okay, thanks!

Sounds good! Thanks for the update.

easwars

What are your thoughts about making this an opt-in feature, i.e, we could have a dial option which the user would have to set for them to get this behavior. If the dial option is not set, the user would get the old behavior.

At a later point in time, we could get rid of the dial option completely and make this the default behavior, or continue to have the dial option (but make the use of the sync.Pool the default one, and the user would have to use the dial option to explicitly disable it).

internal/transport/http_util.go

easwars · 2023-06-27T16:18:41Z

internal/transport/http_util.go

+		}
+		writeBufferPoolMap[size] = pool
+	}
+	writeBufferMutex.Unlock()


When do entries from this map ever get deleted? Can this map grow in a unbounded fashion?

s-matyukevich · 2023-06-27T17:33:18Z

What are your thoughts about making this an opt-in feature, i.e, we could have a dial option which the user would have to set for them to get this behavior. If the dial option is not set, the user would get the old behavior.

At a later point in time, we could get rid of the dial option completely and make this the default behavior, or continue to have the dial option (but make the use of the sync.Pool the default one, and the user would have to use the dial option to explicitly disable it).

Sure, will do that. If we are not enabling this by default, what do you think if I add a similar fix for the read buffer in the same PR? I can use a single option to enable both.

easwars · 2023-06-27T20:36:54Z

Sure, will do that. If we are not enabling this by default, what do you think if I add a similar fix for the read buffer in the same PR? I can use a single option to enable both.

Yes, totally. And if we are going to have it for the server as well, we need a ServerOption as well. Thanks.

easwars

There is a small vet failure with the docstring not starting with the name of the function. Could you please take care of that. Thanks.

dfawley

LGTM, thanks! Just a couple minor things before we can merge it.

server.go

dialoptions.go

internal/transport/http_util.go

ukai · 2023-09-08T03:01:28Z

https://pkg.go.dev/google.golang.org/grpc#WithSharedWriteBuffer

 If this option is set to true every connection will release the buffer after flushing the data on the wire.

not release, but reuse ?

s-matyukevich · 2023-09-08T04:27:08Z

https://pkg.go.dev/google.golang.org/grpc#WithSharedWriteBuffer
 If this option is set to true every connection will release the buffer after flushing the data on the wire.
not release, but reuse ?

release, so it can be reused by other connections.

Fixes siderolabs#7576 See grpc/grpc-go#6309 Signed-off-by: Andrey Smirnov <[email protected]>

raghav-stripe · 2024-02-14T20:09:20Z

internal/transport/http_util.go

+func getWriteBufferPool(writeBufferSize int) *sync.Pool {
+	writeBufferMutex.Lock()
+	defer writeBufferMutex.Unlock()
+	size := writeBufferSize * 2


What is the significance of the * 2 here?

IIRC this matches the legacy behavior, which always doubled for reasons nobody knows but we don't want to change now due to it affecting existing applications.

That's fair. Do you mind expanding on the legacy behavior you are describing? It seems like if there is no shared buffer pool, the buffer is allocated per the input size. It only seems to be in the case of the buffer pool that we're allocating 2x the size (and this seems to be the PR the write buffer pool was added)?

You may be right.. The old code multiplied by 2 when doing this allocation, so it seems we have changed the behavior for legacy users already:

https://github.com/grpc/grpc-go/blame/d524b409462c601ef3f05a7e1fba19755a337c77/internal/transport/http_util.go#L321

Yeah - it just felt somewhat bizarre because this is the very PR that also removed that same * 2 in the default path (see line 321 above), but added the * 2 for the buffer pool path.

If you think it's worth updating, I created #6983

s-matyukevich added 9 commits May 15, 2023 11:56

Release write buffer after Flush() in transport

0f1058d

Use sync.Pool

63f4ba3

Use interface{} instead of any

4faff38

use pool map

f594b29

Merge branch 'master' of github.com:grpc/grpc-go into dynamic-buffer-v1

ec75cb1

fix flush()

95bde71

Merge branch 'master' of github.com:grpc/grpc-go into dynamic-buffer-v1

a243c24

Fix linter error

a7ededc

make variable names more descripptive

16345bb

s-matyukevich commented May 23, 2023

View reviewed changes

internal/transport/http_util.go Show resolved Hide resolved

Merge branch 'master' of github.com:grpc/grpc-go into dynamic-buffer-v1

c8036d0

arvindbr8 self-assigned this May 23, 2023

Merge branch 'master' of github.com:grpc/grpc-go into dynamic-buffer-v1

09407f4

s-matyukevich mentioned this pull request May 24, 2023

benchmark: Add sleepBetweenRPCs and connections parameters #6299

Merged

dfawley added this to the 1.57 Release milestone Jun 2, 2023

s-matyukevich mentioned this pull request Jun 7, 2023

Fix preloader mode in benchmarks #6359

Merged

easwars added the Type: Performance Performance improvements (CPU, network, memory, etc) label Jun 27, 2023

easwars reviewed Jun 27, 2023

View reviewed changes

easwars assigned s-matyukevich Jun 27, 2023

easwars added the Status: Requires Reporter Clarification label Jun 27, 2023

easwars removed the Status: Requires Reporter Clarification label Jun 27, 2023

easwars self-assigned this Jun 27, 2023

Merge branch 'master' of github.com:grpc/grpc-go into dynamic-buffer-v1

0b42642

s-matyukevich added 2 commits July 12, 2023 11:32

Merge branch 'master' of github.com:grpc/grpc-go into dynamic-buffer-v1

d274f35

address review comments

00e8de7

easwars approved these changes Jul 12, 2023

View reviewed changes

easwars unassigned easwars and s-matyukevich Jul 12, 2023

fix linter error

9aeb148

easwars modified the milestones: 1.57 Release, 1.58 Release Jul 14, 2023

arvindbr8 requested a review from dfawley July 18, 2023 17:05

atollena mentioned this pull request Jul 20, 2023

how can limit the grpc connection numbers? #6460

Closed

dfawley requested changes Jul 20, 2023

View reviewed changes

server.go Show resolved Hide resolved

dialoptions.go Show resolved Hide resolved

internal/transport/http_util.go Outdated Show resolved Hide resolved

dfawley assigned s-matyukevich and unassigned dfawley Jul 20, 2023

dfawley added the Status: Requires Reporter Clarification label Jul 20, 2023

s-matyukevich added 2 commits July 20, 2023 14:09

Merge branch 'master' of github.com:grpc/grpc-go into dynamic-buffer-v1

195a467

Add Experimental tag, rename flush to flushKeepBuffer

82a400e

dfawley approved these changes Jul 20, 2023

View reviewed changes

dfawley merged commit 9bb44fb into grpc:master Jul 20, 2023
1 check passed

smira added a commit to smira/talos that referenced this pull request Sep 12, 2023

chore: update gRPC library and enable shared write buffers

59c74b2

Fixes siderolabs#7576 See grpc/grpc-go#6309 Signed-off-by: Andrey Smirnov <[email protected]>

smira mentioned this pull request Sep 12, 2023

chore: update gRPC library and enable shared write buffers siderolabs/talos#7744

Merged

whs mentioned this pull request Sep 13, 2023

Upgrade libraries wongnai/xds#8

Merged

smira added a commit to smira/talos that referenced this pull request Sep 13, 2023

chore: update gRPC library and enable shared write buffers

a096f05

Fixes siderolabs#7576 See grpc/grpc-go#6309 Signed-off-by: Andrey Smirnov <[email protected]>

ash2k mentioned this pull request Oct 5, 2023

rpc_util: Reuse memory buffer for receiving message with stat handlers #6660

Closed

olegbespalov mentioned this pull request Jan 18, 2024

Investigate GRPC's [With]SharedWriteBuffer grafana/k6#3548

Open

pablodz mentioned this pull request Nov 3, 2023

Memory leak on transport.newBufWriter #6761

Closed

raghav-stripe reviewed Feb 14, 2024

View reviewed changes

raghav-stripe mentioned this pull request Feb 15, 2024

transport: when using write buffer pooling, use input size instead of size*2 #6983

Merged

github-actions bot locked as resolved and limited conversation to collaborators Aug 14, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Use sync.Pool to share per-connection transport write buffer. #6309

Use sync.Pool to share per-connection transport write buffer. #6309

s-matyukevich commented May 23, 2023 •

edited by dfawley

Loading

s-matyukevich May 23, 2023 •

edited

Loading

easwars Jun 27, 2023

s-matyukevich Jun 27, 2023

easwars Jun 27, 2023

s-matyukevich commented May 26, 2023

s-matyukevich commented May 30, 2023

dfawley commented May 31, 2023

s-matyukevich commented May 31, 2023

easwars left a comment

easwars Jun 27, 2023

s-matyukevich commented Jun 27, 2023

easwars commented Jun 27, 2023 •

edited

Loading

easwars left a comment

dfawley left a comment

ukai commented Sep 8, 2023

s-matyukevich commented Sep 8, 2023

raghav-stripe Feb 14, 2024 •

edited

Loading

dfawley Feb 14, 2024

raghav-stripe Feb 14, 2024

dfawley Feb 14, 2024

raghav-stripe Feb 15, 2024

raghav-stripe Feb 15, 2024

Use sync.Pool to share per-connection transport write buffer. #6309

Use sync.Pool to share per-connection transport write buffer. #6309

Conversation

s-matyukevich commented May 23, 2023 • edited by dfawley Loading

s-matyukevich May 23, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

s-matyukevich commented May 26, 2023

s-matyukevich commented May 30, 2023

dfawley commented May 31, 2023

s-matyukevich commented May 31, 2023

easwars left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

s-matyukevich commented Jun 27, 2023

easwars commented Jun 27, 2023 • edited Loading

easwars left a comment

Choose a reason for hiding this comment

dfawley left a comment

Choose a reason for hiding this comment

ukai commented Sep 8, 2023

s-matyukevich commented Sep 8, 2023

raghav-stripe Feb 14, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

s-matyukevich commented May 23, 2023 •

edited by dfawley

Loading

s-matyukevich May 23, 2023 •

edited

Loading

easwars commented Jun 27, 2023 •

edited

Loading

raghav-stripe Feb 14, 2024 •

edited

Loading