xpra icon
Bug tracker and wiki

Opened 5 years ago

Last modified 4 months ago

#619 new enhancement

better TCP_NODELAY handling: only use it when it is useful

Reported by: Antoine Martin Owned by: Jonathan Anthony
Priority: critical Milestone: 2.3
Component: core Version: 0.12.x
Keywords: Cc:

Description (last modified by Antoine Martin)

Follow up from #514: at present we enable TCP_NODELAY globally which is a bit wasteful.

It ensures that packets go out as soon as we queue them, but when the packets contain large-ish binary data this means that the binary data and the actual xpra packet structure are likely to travel in separate TCP-level packets.

It would be better to only enable TCP_NODELAY when aggregating packets is not helping: when we have no more data to send or when the output buffer is full. As per: Is there a way to flush a POSIX socket? and this answer:
What I do is enable Nagle, write as many bytes (using non-blocking I/O) as I can to the socket (i.e. until I run out of bytes to send, or the send() call returns EWOULDBLOCK, whichever comes first), and then disable Nagle again. This seems to work well (i.e. I get low latency AND full-size packets where possible)

Good read: The Caveats of TCP_NODELAY

Attachments (2)

tcp.zip (277.2 KB) - added by J. Max Mena 6 months ago.
TCP output data
619tcp.zip (500.8 KB) - added by J. Max Mena 5 months ago.

Download all attachments as: .zip

Change History (17)

comment:5 Changed 3 years ago by Antoine Martin

See also: #1211.

comment:8 Changed 2 years ago by Antoine Martin

See also #639, #999, #401, #540, #417

comment:11 Changed 17 months ago by Antoine Martin

Description: modified (diff)
Priority: majorcritical

I'm seeing some weird behaviour with win32 clients trying to improve #999 and detecting late acks.

The network layer's source_has_more function may not be the right place to set and unset NODELAY because lots of small screen updates can take under a millisecond to compress, which is still slower than it takes the network layer to send them...
Maybe we need to pass the screen update flush attribute down to the network layer too.

comment:12 Changed 17 months ago by Antoine Martin

Owner: changed from Antoine Martin to J. Max Mena

Done in r18149 + r18150. Implementation notes:

  • we add an attribute to the packet source functions so the sender can specify if there will be more packets coming (also used if we know the packet is not urgent and can be delayed by the network layer)
  • we no longer set NODELAY globally on startup, we change (if needed) for each packet
  • in the case of screen updates, we re-use the "flush" attribute (set to a non-zero value when we know that there are more screen update packets coming immediately after)
  • some other packets set the flush flag because they need to be sent in a timely manner (ie: ping) or because they affect the user experience (ie: lost-window)
  • TCP_SOCKET_NODELAY is now XPRA_SOCKET_NODELAY as it applies to more than just TCP sockets
  • we expose the "nodelay" attribute, it should always match the TCP sockopt
  • important bandwidth management updates in ticket:999#comment:44

As of r18151, we can use XPRA_SOCKET_NODELAY to overrule the automatic settings:

  • XPRA_SOCKET_NODELAY=1 should be more or less equivalent to what we had before
  • XPRA_SOCKET_NODELAY=0 never sets NODELAY

TODO:

  • deal with other socket types - websockify does its own thing with a global flag, can we override it?
  • verify RFB and UDP still work after protocol layer changes (and proxy?)
  • maybe keep NODELAY off when there is congestion and we are batching? (not sure this is needed or helpful)
  • maybe audio packets need to be flushed quicker? (does this change cause audio latency to go up?)

@maxmylyn: this ticket is tailor made for the automated tests - we want to compare before and after to see if this helps, especially under bandwidth constrained conditions. (only slight problem is, there is a bug I'm working on which causes congestion detection to kick in too early, and the counter measures are too aggressive, causing the framerate to drop..)

Last edited 17 months ago by Antoine Martin (previous) (diff)

comment:13 Changed 17 months ago by J. Max Mena

I did some maintenance this morning on the test box - it's been acting up again. For starters, I've split the TC tests into two different files - one runs just the packet loss/delay tests, the other runs bandwidth limited cases. That way if something goes wrong with either test suite, we don't lost all the data.

However there's a separate problem that I'm going to look at now - seemingly at random the tests fail to stop the server properly, at which point all following tests fail consistently because they're unable to spin up a server on the display that's still in use. A simple xpra stop clears the server, so I'll sprinkle that command in to the scripts between test suites....but I'd like to figure out why I need to do that.

I'll hold on to this ticket for a few days so we can gather more data.

comment:14 Changed 15 months ago by Antoine Martin

Milestone: 0.152.3

comment:16 Changed 6 months ago by J. Max Mena

As I mentioned in #1840, my mathematician is unavailable for a little bit, and he was working on the new charts, so I'm going to post the raw data I have so far.

Changed 6 months ago by J. Max Mena

Attachment: tcp.zip added

TCP output data

comment:17 Changed 5 months ago by Antoine Martin

TCP output data

It's not clear what command lines were used for each run, as there are 3 possible values for XPRA_SOCKET_NODELAY (0, 1, unset)

It also doesn't look like this was being tested with any bandwidth constraints?
All encodings are tested in there, which is going to make the data very noisy. I would focus on one setting ("auto" or "rgb"), and only one test (one that generates lots of small-ish screen updates - maybe simulate console user, or gtkperf) and maybe vary only the throttling - if anything.

Last edited 5 months ago by Antoine Martin (previous) (diff)

comment:18 Changed 5 months ago by J. Max Mena

It's not clear what command lines were used for each run, as there are 3 possible values for XPRA_SOCKET_NODELAY (0, 1, unset)

For these tests, I have it run a series of three test runs. There are two sets - the ones with the prefix nodelay_ have XPRA_SOCKET_NODELAY=1 set, and delay_ has XPRA_SOCKET_NODELAY=0 set.

It also doesn't look like this was being tested with any bandwidth constraints?

I'll copy some of the Bash script from when I was running daily tests with bandwidth constraints - 25, 16, and 8 megabits should be a good enough starting point. I've changed the config file for this test run to only use rgb, and it only runs the two console tests and gtkperf - since I'm not hardware limited anymore, it doesn't hurt to run extra tests since I'll usually have a spare machine or two in case I need it for something else.

Changed 5 months ago by J. Max Mena

Attachment: 619tcp.zip added

comment:19 Changed 5 months ago by J. Max Mena

I've posted a new set of data. This time I ran tc qdisc add dev lo root netem rate 25mbit between test runs with XPRA_SOCKET_NODELAY set to 0 and 1, with 25mbit, 16mbit, and 8mbit bandwidth constraints.

Unfortunately most of that data is useless. It looks like the IP Tables command didn't work (EDIT: This one's on me - I left USE_IPTABLES disabled in the config), so all the packet accounting didn't get recorded, and as mentioned in #1840 quite a few of the other columns are missing as well. I'll try to figure out why that's the case, but I'm posting it anyways since all the data is ~500KB.

Last edited 5 months ago by J. Max Mena (previous) (diff)

comment:20 Changed 5 months ago by Antoine Martin

.. as mentioned in #1840 quite a few of the other columns are missing as well ..

As mentioned in #1840, the breakage is likely quite recent whereas this ticket is a year old.
Why not test with a version that does record the data we want?

comment:21 Changed 5 months ago by Antoine Martin

r21493 waits until after we have sent the last chunk before enabling NODELAY.
I believe that's correct and the kernel will then flush the socket, but it would be much better to verify that with test data.

comment:22 Changed 5 months ago by Antoine Martin

r21495: also disable NODELAY for multiple chunks (doh)

comment:23 Changed 5 months ago by Antoine Martin

See also #2130

comment:24 Changed 4 months ago by Antoine Martin

Owner: changed from J. Max Mena to Jonathan Anthony
Note: See TracTickets for help on using tickets.