Re^7: Multithreaded Server Crashes under heavy load

[BrowserUk]

in reply to Re^6: Multithreaded Server Crashes under heavy load
in thread Multithreaded Server Crashes under heavy load

It is all over after the first 3 minutes. Its not a "hang". It's just running really slowly because it cannot get new sockets for connections.

You are creating huge numbers of connections -- there are 6,897 open sockets when your program runs, presumably still timing out from previous runs -- but rather than cleanly shutting those connections down, they are going into TIME_WAIT state and then your server has to wait for one of them to time out (900 seconds or some such) before it can establish a new connection.

MST                          Elapsed Time           Working Set       
+      Established      Reset
                    Processor Time          Thread Count            Ac
+tive          Passive

18:45:38.609        0        0              0               0       68
+97    6       6062    505
18:45:48.609                 5.9687         1        12566528       68
+97    5       6062    505
18:45:58.609        16.0937 15.96875        5        40304640       68
+97    8       6081    505
18:46:08.609        23.4375 25.96875        13       78077952       68
+97    60      6166    505
18:46:18.609        38.9062 35.96875        12      109629440       68
+98    15      6185    505
18:46:28.625        52.1060 45.984375       8        94683136       68
+98    11      6185    505
18:46:38.640        59.5943 56              2        21966848       68
+98    5       6185    505
18:46:48.656        15.6006 66.015625       11       58703872       68
+98    47      6259    505
18:46:58.656        27.5000 76.015625       41      283103232       68
+98    69      6309    506
18:47:08.671        17.0046 86.03125        41      475500544       68
+98    69      6309    506
18:47:18.671        14.8437 96.03125        41      536985600       68
+98    69      6309    506
18:47:28.671        0       106.03125       41      536985600       68
+98    69      6309    506
18:47:38.671        0       116.03125       41      536985600       68
+99    69      6309    506
18:47:48.671        0       126.03125       41      536985600       68
+99    119     6359    506
18:47:58.671        0       136.03125       41      536985600       68
+99    192     6432    506
18:48:08.671        0       146.03125       41      536985600       68
+99    192     6432    506
18:48:18.671        0       156.03125       41      536985600       68
+99    192     6432    506
18:48:28.671        0       166.03125       41      536985600       68
+99    193     6433    506
18:48:38.671        0       176.03125       41      536985600       68
+99    193     6433    506
18:48:48.671        0       186.03125       41      536985600       68
+99    193     6433    506
18:48:58.671        0       196.03125       41      536985600       68
+99    191     6433    508
18:49:08.671        0       206.03125       41      536985600       68
+99    192     6434    508
18:49:18.671        0       216.03125       41      536985600       68
+99    192     6434    508
18:49:28.671        0       226.03125       41      536985600       68
+99    241     6535    508
18:49:38.671        0       236.03125       41      536985600       68
+99    241     6558    508
18:49:48.671        0       246.03125       41      536985600       68
+99    240     6559    509

(If you are going to throw another set of data at us, how about you expend a little energy to make the csv data readable :)

There is either:

• something wrong with the architecture of your server;

  (I haven't spotted it yet, but without running and tracing it can be hard to spot);

  The debug log -- had it not been empty - might have helped.

• or your clients are not closing their ends of the connections properly;

  Basically, you are in a "dead man's shoes" situation, where your server cannot establish (or probably even accept) a new connection, until one of the existing dying-but-still-to-finally-die connections times out.

Before you do another run, you need to clean up any existing connections. There is the netsh command that allows you to reset at various levels -- winsock; interface; ipv4; tcp etc. -- but perhaps the simplest is to just reboot the machine.

You need to work out what is causing the connections to 'linger'. You appear to be using shutdown correctly -- at the server end at least -- and closing the filehandles; but something is preventing them from being reused immediately, despite your ReuseAddr setting on the listener.

Not a solution, but maybe it will give you some clues about where to start looking and how.

[Anonymous Monk]

I think  net stop "dhcp client" && net start "dhcp client" ought to do that quick :) if you're using dhcp that is

[rmahin]

Hey sorry about the CSV! I just opened it up in Libre Office as a comma delimited file, and it looked good. BUT! Good news. Think you're definitely right that I'm not closing connections properly. Think this is caused by commands that need to open other file handles, and not handling any errors correctly. For instance, I think the only command I included that has that behavior is the PUT command, which does  open( PUTFILE, ">$outfile" ) or threads->exit;. The threads->exit is clearly getting called, leaving the socket open. I fixed all those and added more messages to see when exactly it was happening, and it definitely seems to be working better. Have on more part of the code to look at using a subroutine that does not appear to be thread friendly, but will update on progress once I've determined if I solved it or not. Thanks for the direction!

[BrowserUk]

which does open( PUTFILE, ">$outfile" ) or threads->exit;. The threads->exit is clearly getting called, leaving the socket open.

FWIW: I have written a crap load of threaded perl code and never had occasion to use thread->exit;. It is IMO redundant and dangerous.

I would code that line as simply:

open( PUTFILE, ">$outfile" ) or return;

That way, all the normal perl cleanup will take place before the thread function returns and the thread terminates.

Try it, it just might sort out a lot of your problems!

[rmahin]

Hey sorry for the late reply, but I think everything is working great now! I essentially did what you suggested so looks like that was what was killing me. Been running for a week and no problems. Thanks very much for the help.