SCTP throughput ~8x slower than TCP when the connection has latency

[iamcalledrob]

Your environment.

• Version: sctp v1.8.2
• Other Information: tested between macOS<–>macOS and macOS<–>debian bullseye

What did you do?

Test setup:

• Throughput tested with @enobufs's sctptest, dependencies updated to use the latest version of pion/sctp
• Latency controlled using macOS "Network Link Conditioner"
• Tested between separate physical devices connected by gigabit ethernet

I ran sctp-test in both TCP and SCTP (sctp-test refers to this as "udp") modes and compared the throughput with varying levels of latency:

• For TCP: ./sctptest -network tcp -s <addr> [client] and ./sctptest -network tcp [server]
• For SCTP: ./sctptest -network udp -s <addr> [client] and ./sctptest -network udp [server]

What did you expect?

I expected the TCP and SCTP throughput to be in the same ballpark.

What happened?

SCTP overwhelmingly underperformed TCP. Approximate performance over gigabit ethernet:

Latency	TCP	SCTP
2ms	850 Mbps	850 Mbps
20ms	800 Mbps	200 Mbps
50ms	600 Mbps	85 Mbps
100ms	320 Mbps	45 Mbps
300ms	105 Mbps	17 Mbps

Background

I ran into this performance issue when trying to diagnose why DataChannel performance was very slow when using a TURN server. My TURN server is about 100ms away, and I was seeing very poor performance (< 10 Mbps) relative to the speed of the internet connection.

SCTP.throughput.and.latency.mp4

[iamcalledrob]

I found this issue from a few years back which looks potentially related? #62

[Sean-Der]

Fantastic testing @iamcalledrob thanks for reporting!

I don't know the latest but @enobufs mentioned he was looking at adding RACK to the implementation. Maybe some other things exist that we can do in the short term? I defer to him on what the best thing to do is :)

[enobufs]

@iamcalledrob Thanks for the detailed report. As I mentioned in #62, a modern TCP implementation automatically adjust receive buffer size, while pion/sctp uses a fixed buffer size, which is currently set to 1024KB. The receive buffer size sets the hard limit for the amount of data inflight (or the un-ack'd segments). It appears the throughput you observed is the result of this limit. I think pion/sctp should implement this auto tuning feature. Let me dig into more information about its algorithm.

[iamcalledrob]

@Sean-Der, @enobufs Thank you both for looking into this!

@enobufs regarding the buffer size, I was curious about that as well and decided to do some more testing!

I benchmarked the effect of changing the buffer size via your sctp-test utility—i.e. sctp-test -network udp -b N. I believe this sets both the read and write buffer sizes to 'N bytes' on the SCTP association as well as the underlying UDP conn.

Here's what I saw with round-trip latency of 300ms

Recv buffer size	Throughput
16kb	0.5 Mbps
64kb	13 Mbps
256kb	17 Mbps
1mb	17 Mbps
4mb	17 Mbps

As you can see, the throughput maxes out with a buffer size of 256kb—smaller than the default 1mb! Increasing the buffer size beyond 256kb has no effect.

I'm curious about this—I would have assumed that manually setting a larger buffer would have improved throughput in this particular test, but that has not been the case.

I'm happy to help with getting to the bottom of this, though I have to admit I'm not coming from a place of expertise here so pointers are very helpful :)

I've attached trace logs for both the client and server in case it's helpful context.

client.log
server.log

[enobufs]

That indicates there would be another limiting factor. From my quick glance, the sender's congestion window kept growing during the 10 sec period but it appears too slow - smell of a bug... I may be wrong. Let me look at this later more carefully.

[enobufs]

I just found a bug in sctptest. The value with the -b option for the server does not propagate down to pion/sctp :( .

What's going on is, the client side should have a ssthresh (slow start threshold) which should initially be set to the receive buffer window advertised by the peer. This value did not change even though I was increasing the value with -b, then found this bug. I will fix it shortly.

[enobufs]

After fixing the above in sctptest, I was still seeing the throughput was still capped at 1MB receive buffer size. (increasing the size beyond 1MB had no effect).

sctptest RTT=300ms

From the trance log, I noticed that even though there's more room in the congestion window, the inflightQueue was not fully utilized. Then I realized, this was because sctptest has maxBufferedAmount set to 1MB (threshold being 512KB, both hardcoded), which was effectively limiting the send buffer size to 1MB. This is why the throughput did not grow beyond 1MB of receive buffer size.

There was another related issue. The bufferedAmount in the current pion/sctp includes the number of bytes inflight. Looking at the FireFox's implementation, it does not include the size inflight (the size written to usrsctp).

After fixing the all above, now I see the throughput grows proportionally to the size of buffers as shown below:

Other related metrics:

I will write some unit tests then submit a pull-request.

Once all these fixes land, we should consider whether we should dynamically adjust the receive buffer size.

[enobufs]

Following updates have been made:

• go-rudp v0.0.3
• sctp-test v0.0.2

TODO:

• pion/sctp: exclude inflight size from bufferedAmount

[iamcalledrob]

Thank you for looking into this! I'm excited to see what the throughput looks like when inflight size is excluded :)

[iamcalledrob]

With your updated version of go-rudp/sctp-test (passing buffer size into pion/sctp assoc), I see dramatically better throughput.

For the curious:

RTT	Recv buffer size	Throughput
20ms	1MB	370Mbps
20ms	2MB	700Mbps
20ms	4MB	880Mbps
---	---	---
50ms	1MB	150Mbps
50ms	2MB	300Mbps
50ms	4MB	590Mbps
50ms	8MB	800Mbps
---	---	---
300ms	1MB	25Mbps
300ms	2MB	55Mbps
300ms	4MB	100Mbps
300ms	8MB	200Mbps
300ms	24MB	310Mbps
300ms	32MB+	(Unreliable)

This is faster than TCP.
To repro these tests, I had to make a small fix to cap conn.SetReadBuffer/SetWriteBuffer to ~5mb for mac (the OS UDP limit), while allowing the sctp buffer to be the specified size

Very excited for the use-cases this unlocks for pion/sctp!

[Sean-Der]

That's so exciting! Nice work @iamcalledrob and @enobufs :)

We should add this to the SettingEngine today and publish the findings!

[enobufs]

@Sean-Der I think it is a brilliant idea to allow changing the send/recv buffer size from the SettingEngine as I think it may take sometime for us to implement auto buffer size tuning.