Changes between Initial Version and Version 1 of PlasticSlices/Continuation/Round4

Timestamp:

03/07/12 10:49:16 (12 years ago)

Author:

Josh Smift

Comment:

--

PlasticSlices/Continuation/Round4

@@ +1 @@
+[[PageOutline]]
+
+= Continuation Round 4 =
+
+In this round, we ran plastic-105 and plastic-106 for about three hours, with only GT and Wisconsin, to try to track down some issues with performance in Round 3 that seemed related to GT. In particular, we wanted to find out whether there was a specific problem with plnode1.
+
+The raw logs are at http://www.gpolab.bbn.com/plastic-slices/continuation/round-4/. 
+
+= plastic-105 =
+
+[ggw:PlasticSlices/Experiments#SteadyPerfTCP SteadyPerf TCP], using port=5105, time=518400, and this table of client/server pairs:
+
+|| '''client'''                   || '''server'''                   || '''server address''' ||
+|| plnode1.cip.gatech.edu         || wings-openflow-2.wail.wisc.edu || server=10.42.105.95  ||
+|| plnode2.cip.gatech.edu         || wings-openflow-3.wail.wisc.edu || server=10.42.105.96  ||
+
+== One-time prep commands run on each client and server ==
+
+{{{
+sudo yum -y install iperf
+}}}
+
+== Commands run on each server ==
+
+{{{
+server=<ipaddr>
+nice -n 19 iperf -B $server -p 5105 -s -i 1
+}}}
+
+== Commands run on each client ==
+
+{{{
+server=<ipaddr>
+nice -n 19 iperf -c $server -p 5105 -t 518400
+}}}
+
+== Results ==
+
+Generated with
+
+{{{
+subnet=105
+for host in $(awk 'NR%2==1' ~/plastic-slices/logins/logins-plastic-$subnet.txt | sed -r -e 's/.+@//') ; do echo -e "$host:\n\n{{{" ; grep -A 5 -B 1 "Client connecting" pgenigpolabbbncom_plastic$subnet\@$host.log ; echo -e "}}}\n" ; done
+}}}
+
+and then edited slightly to remove artifacts (like control characters, my prompt, etc).
+
+plnode1.cip.gatech.edu:
+
+{{{
+------------------------------------------------------------
+Client connecting to 10.42.105.95, TCP port 5105
+TCP window size: 16.0 KByte (default)
+------------------------------------------------------------
+[  3] local 10.42.105.100 port 59778 connected with 10.42.105.95 port 5105
+[  3]  0.0-13761.8 sec    405 MBytes    247 Kbits/sec
+}}}
+
+plnode2.cip.gatech.edu:
+
+{{{
+------------------------------------------------------------
+Client connecting to 10.42.105.96, TCP port 5105
+TCP window size: 16.0 KByte (default)
+------------------------------------------------------------
+[  3] local 10.42.105.101 port 58242 connected with 10.42.105.96 port 5105
+[  3]  0.0-13768.4 sec    482 MBytes    294 Kbits/sec
+}}}
+
+== Analysis ==
+
+Less than 300 Kbits/sec is very low.
+
+Looking at the server logs shows times when the rate jumped up to one or two Mbits/sec, before dropping back down to double-digt Kbits/sec.
+
+The results are similar for the two plnodes, so there probably isn't a problem with just one or the other.
+
+= plastic-106 =
+
+[ggw:PlasticSlices/Experiments#SteadyPerfUDP SteadyPerf UDP], using port=5106, time=518400, rate=30, and this table of client/server pairs:
+
+|| '''client'''                   || '''server'''                   || '''server address''' ||
+|| plnode1.cip.gatech.edu         || wings-openflow-3.wail.wisc.edu || server=10.42.105.96  ||
+|| plnode2.cip.gatech.edu         || wings-openflow-2.wail.wisc.edu || server=10.42.105.95  ||
+
+== One-time prep commands run on each client and server ==
+
+{{{
+sudo yum -y install iperf
+}}}
+
+== Commands run on each server ==
+
+{{{
+server=<ipaddr>
+nice -n 19 iperf -u -B $server -p 5106 -s -i 1
+}}}
+
+== Commands run on each client ==
+
+{{{
+server=<ipaddr>
+nice -n 19 iperf -u -c $server -p 5106 -t 518400 -b 30M
+}}}
+
+== Results ==
+
+Generated with
+
+{{{
+subnet=106
+for host in $(awk 'NR%2==1' ~/plastic-slices/logins/logins-plastic-$subnet.txt | sed -r -e 's/.+@//') ; do echo -e "$host:\n\n{{{" ; grep -A 10 -B 1 "Client connecting" pgenigpolabbbncom_plastic$subnet\@$host.log ; echo -e "}}}\n" ; done
+}}}
+
+and then edited slightly to remove artifacts (like control characters, my prompt, etc).
+
+plnode1.cip.gatech.edu:
+
+{{{
+------------------------------------------------------------
+Client connecting to 10.42.105.96, UDP port 5106
+Sending 1470 byte datagrams
+UDP buffer size:   122 KByte (default)
+------------------------------------------------------------
+[  3] local 10.42.105.100 port 53151 connected with 10.42.105.96 port 5106
+[  3]  0.0-13755.2 sec  48.0 GBytes  30.0 Mbits/sec
+[  3] Sent 35057619 datagrams
+[  3] Server Report:
+[  3]  0.0-13754.8 sec  44.2 GBytes  27.6 Mbits/sec  0.078 ms 2767576/35057619 (7.9%)
+[  3]  0.0-13754.8 sec  252 datagrams received out-of-order
+}}}
+
+plnode2.cip.gatech.edu:
+
+{{{
+------------------------------------------------------------
+Client connecting to 10.42.105.95, UDP port 5106
+Sending 1470 byte datagrams
+UDP buffer size:   122 KByte (default)
+------------------------------------------------------------
+[  3] local 10.42.105.101 port 39327 connected with 10.42.105.95 port 5106
+[  3]  0.0-13756.4 sec  48.0 GBytes  30.0 Mbits/sec
+[  3] Sent 35068319 datagrams
+[  3] Server Report:
+[  3]  0.0-13756.0 sec  44.4 GBytes  27.7 Mbits/sec  0.252 ms 2666045/35068319 (7.6%)
+[  3]  0.0-13756.0 sec  319 datagrams received out-of-order
+}}}
+
+== Analysis ==
+
+Similar packet loss to Round 3, where we saw 7.5% loss.
+
+The results are similar for the two plnodes, so there probably isn't a problem with just one or the other.