Hide Forgot
Description of problem: Using Solarflare 10Gb card, MRG-M perf test fails when testing packet sizes grater than 512MB Version-Release number of selected component (if applicable): 2.6.33.9-rt31.74.el6rt.x86_64 How reproducible: consistently Steps to Reproduce: 1. boot RT kernel 2. Start qpidd broker one 1 node 3./usr/sbin/qpidd --data-dir /var/lib/qpidd --auth no -m no --log-enable=info+ Actual results: Testing runs in a loop for 100 times, sometime into the run get the following error on the broker node 2011-08-18 13:56:29 error Error reading socket: Bad address(14) 2011-08-18 14:00:20 notice Shut down There are no messages relating to this issue in dmesg or /var/log/messages. Expected results: Should be able to run a complete suite of testing with various packet sizes ranging from 8 Bytes to 32K. This can be completed using non-RT kernel. Also, do not see this issue with Mellanox 10Gb cards. Additional info:
The packet size should state to be 512K or greater
Which non-RT kernel was used?
And what is the client used? Some standard test found in the qpidd packages? What command line was used with it?
On the broker side, /usr/sbin/qpidd --data-dir /var/lib/qpidd --auth no -m no --log-enable=info+ Client side [root@perf46 ~]# cat run_amqp_perf_perf48.sh #!/bin/bash for i in 8 16 32 64 128 256 512 1024 2048 4096 8192 16384 32768; #for i in 1024 2048 4096 8192 16384 32768; #for i in 32768; do echo $i; /root/testNnics.sh 1 100000 $i 192.168.10.48; /root/testNnics.sh 1 100000 $i 192.168.10.48; /root/testNnics.sh 1 100000 $i 192.168.10.48; /root/testNnics.sh 1 100000 $i 192.168.10.48; /root/testNnics.sh 1 100000 $i 192.168.10.48; /root/testNnics.sh 1 100000 $i 192.168.10.48; /root/testNnics.sh 1 100000 $i 192.168.10.48; /root/testNnics.sh 1 100000 $i 192.168.10.48; /root/testNnics.sh 1 100000 $i 192.168.10.48; . . . testNnics.sh #!/bin/sh # Usage: usage() { echo "Usage: $0 pubsubs_per_nic count size nic1 [ nic2 ... ]" echo "Uses $PERFTEST if set in env." exit 1 } test $# -lt 4 && usage pubsubs_per_nic=$1; shift count=$1; shift size=$1; shift nclient=`expr $pubsubs_per_nic \* $#` #echo "Run $nclient pubs & subs with $pubsubs_per_nic on each NIC: $* " test "`expr $count / $nclient`" = 0 || { count=`expr $count + $nclient - \( $count % $nclient \)` # echo "WARNING: Adjusted count to $count, nearest multiple of client count $nclient" } perftest=${PERFTEST:-/usr/bin/qpid-perftest --tcp-nodelay} #perftest=${PERFTEST:-/usr/bin/qpid-perftest -P rdma} args="--nsubs $nclient --npubs $nclient --count $count --size $size --qt 16 --summary" # Can use any nic for --setup and --control clients nic1=$1 $perftest --setup -b $nic1 $args --bounds-multiplier 102400 # NB: not in background. # Split pubsubs_per_nic pubs & subs for each nics. while [ -n "$*" ]; do nic=$1; shift for (( i = 0 ; i < pubsubs_per_nic ; ++i )) ; do $perftest --publish -b $nic $args & $perftest --subscribe -b $nic $args & done done $perftest --control -b $nic1 $args Tried reducing the number of queues (qt) to default of 4 and still see the same issues Tom
With the 2.6.33.9-rt31.66.el6rt.x86_64 still see the same issue [root@perf48 ~]# /usr/sbin/qpidd --data-dir /var/lib/qpidd --auth no -m no 2011-08-18 16:06:48 notice SASL disabled: No Authentication Performed 2011-08-18 16:06:48 notice Listening on TCP port 5672 2011-08-18 16:06:48 notice Broker running 2011-08-18 16:12:10 error Error reading socket: Bad address(14)
This is too old, out of scope. Perhaps the test is worth snagging