Changes between Version 1 and Version 2 of PlasticSlices/Continuation/Round6


Ignore:
Timestamp:
03/08/12 09:55:38 (7 years ago)
Author:
Josh Smift
Comment:

--

Legend:

Unmodified
Added
Removed
Modified
  • PlasticSlices/Continuation/Round6

    v1 v2  
    44
    55In this round, I ran plastic-105 and plastic-106 for about N hours, with only GT and Wisconsin, to try to track down some issues with performance in Round 3 that seemed related to GT. This was just a repeat of Round 4.
     6
     7Further testing (not shown here) suggests that the problem is specific to data leaving the GT plnodes' dataplane interfaces: We get good results when we use the control interfaces, and with data leaving a wapgnode at GT, and bad performance even between nodes purely at GT. The only times we see problems are when GT plnodes are sending data, and it doesn't seem to matter at all where they're sending it. GT is investigating.
    68
    79The raw logs are at http://www.gpolab.bbn.com/plastic-slices/continuation/round-6/.
     
    4648and then edited slightly to remove artifacts (like control characters, my prompt, etc).
    4749
    48 (because the summary of the results are in the client logs, which are the odd-numbered lines in the file)
     50plnode1.cip.gatech.edu:
     51
     52{{{
     53------------------------------------------------------------
     54Client connecting to 10.42.105.95, TCP port 5105
     55TCP window size: 16.0 KByte (default)
     56------------------------------------------------------------
     57[  3] local 10.42.105.100 port 54474 connected with 10.42.105.95 port 5105
     58[  3]  0.0-17262.0 sec  1.52 GBytes    755 Kbits/sec
     59}}}
     60
     61plnode2.cip.gatech.edu:
     62
     63{{{
     64------------------------------------------------------------
     65Client connecting to 10.42.105.96, TCP port 5105
     66TCP window size: 16.0 KByte (default)
     67------------------------------------------------------------
     68[  3] local 10.42.105.101 port 38278 connected with 10.42.105.96 port 5105
     69[  3]  0.0-17264.6 sec  1.93 GBytes    958 Kbits/sec
     70}}}
    4971
    5072== Analysis ==
     73
     74These were somewhat higher than previously because there were periods when we weren't running plastic-106; we discovered that the low TCP bandwidth only happened while plastic-106 was running. The server logs show this more clearly.
    5175
    5276= plastic-106 =
     
    89113and then edited slightly to remove artifacts (like control characters, my prompt, etc).
    90114
    91 (because the summary of the results are in the client logs, which are the odd-numbered lines in the file)
     115plnode1.cip.gatech.edu:
     116
     117{{{
     118------------------------------------------------------------
     119Client connecting to 10.42.105.96, UDP port 5106
     120Sending 1470 byte datagrams
     121UDP buffer size:   122 KByte (default)
     122------------------------------------------------------------
     123[  3] local 10.42.105.100 port 48341 connected with 10.42.105.96 port 5106
     124[  3]  0.0-16304.4 sec  56.9 GBytes  30.0 Mbits/sec
     125[  3] Sent 41551881 datagrams
     126[  3] Server Report:
     127[  3]  0.0-16304.1 sec  52.3 GBytes  27.6 Mbits/sec  0.134 ms 3318803/41551881 (8%)
     128[  3]  0.0-16304.1 sec  240 datagrams received out-of-order
     129}}}
     130
     131plnode2.cip.gatech.edu:
     132
     133{{{
     134------------------------------------------------------------
     135Client connecting to 10.42.105.95, UDP port 5106
     136Sending 1470 byte datagrams
     137UDP buffer size:   122 KByte (default)
     138------------------------------------------------------------
     139[  3] local 10.42.105.101 port 40375 connected with 10.42.105.95 port 5106
     140[  3]  0.0-16319.8 sec  57.0 GBytes  30.0 Mbits/sec
     141[  3] Sent 41600427 datagrams
     142[  3] Server Report:
     143[  3]  0.0-16319.3 sec  52.6 GBytes  27.7 Mbits/sec  0.084 ms 3160813/41600427 (7.6%)
     144[  3]  0.0-16319.3 sec  232 datagrams received out-of-order
     145}}}
    92146
    93147== Analysis ==
     148
     149Much the same packet loss as before.