Thread scheduler misbehaviour OR Oprofile bug?
Hi,
I'm using OProfile to check some suspicious behaviour of dpdk-pktgen, and I can see something which troubles me. Either the scheduler lies about core affinity or Oprofile accounts some samples wrongly.
This userspace app runs in threads, which are assigned explicitly to one single core with pthread_setaffinity_np (I've added some debug lines which checks the affinity with pthread_getaffinity_np to make sure that the affinity was correctly set indeed)
These threads run infinite loops, calling different functions. In my example one of these functions (pktgen_main_rxtx_loop) should only be called by the thread on core #1, but in the oprofile results I can see samples from all other cores as well. I've added a sched_getcpu() call to every iteration of that infinite loop to check if it runs anywhere else than core #1, but it seems to be fine. So my conclusion is that either the scheduler doesn't care about thread affinity and even sched_getcpu() can't see that, OR, Oprofile falsely accounts samples to the wrong CPU. Or to the wrong symbol, although I've added "__attribute__ ((noinline))" to every called function to make sure that inlineing doesn't screw up anything. I'm running my tests on Ubuntu Servers with 14.04 with 3.13.0-32 kernel, the CPU is a desktop kind, i5-4570 @ 3.20GHz (no-HT!) with oprofile 0.9.9.
Anyone have a suggestion about what happens?
Regards,
Zoltan
Quoted 24 lines · Zoltan Kiss
Hi,
I'm using OProfile to check some suspicious behaviour of dpdk-pktgen, and I can see something which troubles me. Either the scheduler lies about core affinity or Oprofile accounts some samples wrongly. This userspace app runs in threads, which are assigned explicitly to one single core with
pthread_setaffinity_np(I've added some debug lines which checks the affinity withpthread_getaffinity_npto make sure that the affinity was correctly set indeed)
These threads run infinite loops, calling different functions. In my example one of these functions (pktgen_main_rxtx_loop) should only be called by the thread on core #1, but in the oprofile results I can see samples from all other cores as well. I've added asched_getcpu()call to every iteration of that infinite loop to check if it runs anywhere else than core #1, but it seems to be fine.
So my conclusion is that either the scheduler doesn't care about thread affinity and evensched_getcpu()can't see that, OR, Oprofile falsely accounts samples to the wrong CPU. Or to the wrong symbol, although I've added "__attribute__((noinline))" to every called function to make sure that inlineing doesn't screw up anything.
I'm running my tests on Ubuntu Servers with 14.04 with 3.13.0-32 kernel, the CPU is a desktop kind, i5-4570 @ 3.20GHz (no-HT!) with oprofile 0.9.9. Anyone have a suggestion about what happens?
Quoted 3 lines
Regards,
Zoltan
Hi Zoltan,
Are you using operf or opcontrol to set up the data measurements? It would be good to provide the command lines that you are using to do the data collection with oprofile.
"perf stat <command>" or "perf stat -p <list_of_pids>" can show task migration. Look for the non-zero "cpu-migrations" in the output like in the output below.
$ perf stat -p 29600 ^C Performance counter stats for process id '29600':
4303.761019 task-clock (msec) # 0.305 CPUs utilized [100.00%]
17,980 context-switches # 0.004 M/sec [100.00%]
595 cpu-migrations # 0.138 K/sec [100.00%]
44,212 page-faults # 0.010 M/sec
10,585,868,260 cycles # 2.460 GHz [100.00%]
6,463,554,435 stalled-cycles-frontend # 61.06% frontend cycles idle [100.00%]
<not supported> stalled-cycles-backend
10,400,529,789 instructions # 0.98 insns per cycle
# 0.62 stalled cycles per insn [100.00%]
2,049,813,299 branches # 476.284 M/sec [100.00%]
48,441,881 branch-misses # 2.36% of all branches
14.112614437 seconds time elapsedAs a check you might try using "perf record <command>" or "perf record -p <list_of_pids>". Once data collection is done use "perf report --sort=cpu" to see whether samples were recorded on other processors. That would determine how much the monitored process is beoing scheduled on other cpus.
There is also a systemtap (https://sourceware.org/systemtap/) script cycle_thief (https://sourceware.org/systemtap/examples/#process/cycle_thief.stp) that shows how often tasks are migrated and what other tasks are competing with the pid for time.
These suggestions should help determine whether oprofile is getting the attribution of samples wrong or the scheduler is migrating the tasks.
-Will
Hi,
Quoted 25 lines · William Cohen
Quoted 24 lines · Zoltan Kiss
Hi,
I'm using OProfile to check some suspicious behaviour of dpdk-pktgen, and I can see something which troubles me. Either the scheduler lies about core affinity or Oprofile accounts some samples wrongly. This userspace app runs in threads, which are assigned explicitly to one single core with
pthread_setaffinity_np(I've added some debug lines which checks the affinity withpthread_getaffinity_npto make sure that the affinity was correctly set indeed)
These threads run infinite loops, calling different functions. In my example one of these functions (pktgen_main_rxtx_loop) should only be called by the thread on core #1, but in the oprofile results I can see samples from all other cores as well. I've added asched_getcpu()call to every iteration of that infinite loop to check if it runs anywhere else than core #1, but it seems to be fine.
So my conclusion is that either the scheduler doesn't care about thread affinity and evensched_getcpu()can't see that, OR, Oprofile falsely accounts samples to the wrong CPU. Or to the wrong symbol, although I've added "__attribute__((noinline))" to every called function to make sure that inlineing doesn't screw up anything.
I'm running my tests on Ubuntu Servers with 14.04 with 3.13.0-32 kernel, the CPU is a desktop kind, i5-4570 @ 3.20GHz (no-HT!) with oprofile 0.9.9. Anyone have a suggestion about what happens?
Quoted 7 lines
Quoted 3 lines
Regards,
Zoltan
Hi Zoltan,
Are you using operf or opcontrol to set up the data measurements? It would be good to provide the command lines that you are using to do the data collection with oprofile.
Thanks for your reply! I'm using operf, here are the command sequence I'm using:
rm -rf oprofile_data
operf -scl --event CPU_CLK_UNHALTED:100000:0:1:1 --vmlinux /usr/lib/debug/boot/vmlinux-uname -r & sleep 20
killall -s SIGINT operf
sleep 1
opreport --image-path /local/pktgen-dpdk/app/build/ --symbols|less
I'll check the rest of your suggestions later on.
Zoltan
Quoted 31 lines
"perf stat <command>" or "perf stat -p <
list_of_pids>" can show task migration. Look for the non-zero "cpu-migrations" in the output like in the output below.$ perf stat -p 29600 ^C Performance counter stats for process id '29600':4303.761019 task-clock (msec) # 0.305 CPUs utilized [100.00%] 17,980 context-switches # 0.004 M/sec [100.00%] 595 cpu-migrations # 0.138 K/sec [100.00%] 44,212 page-faults # 0.010 M/sec 10,585,868,260 cycles # 2.460 GHz [100.00%] 6,463,554,435 stalled-cycles-frontend # 61.06% frontend cycles idle [100.00%] <not supported> stalled-cycles-backend 10,400,529,789 instructions # 0.98 insns per cycle # 0.62 stalled cycles per insn [100.00%] 2,049,813,299 branches # 476.284 M/sec [100.00%] 48,441,881 branch-misses # 2.36% of all branches 14.112614437 seconds time elapsedAs a check you might try using "perf record <command>" or "perf record -p <
list_of_pids>". Once data collection is done use "perf report --sort=cpu" to see whether samples were recorded on other processors. That would determine how much the monitored process is beoing scheduled on other cpus.There is also a systemtap (https://sourceware.org/systemtap/) script
cycle_thief(https://sourceware.org/systemtap/examples/#process/cycle_thief.stp) that shows how often tasks are migrated and what other tasks are competing with the pid for time.These suggestions should help determine whether oprofile is getting the attribution of samples wrong or the scheduler is migrating the tasks.
-Will
Quoted 39 lines · William Cohen
Quoted 24 lines · Zoltan Kiss
Hi,
I'm using OProfile to check some suspicious behaviour of dpdk-pktgen, and I can see something which troubles me. Either the scheduler lies about core affinity or Oprofile accounts some samples wrongly. This userspace app runs in threads, which are assigned explicitly to one single core with
pthread_setaffinity_np(I've added some debug lines which checks the affinity withpthread_getaffinity_npto make sure that the affinity was correctly set indeed)
These threads run infinite loops, calling different functions. In my example one of these functions (pktgen_main_rxtx_loop) should only be called by the thread on core #1, but in the oprofile results I can see samples from all other cores as well. I've added asched_getcpu()call to every iteration of that infinite loop to check if it runs anywhere else than core #1, but it seems to be fine.
So my conclusion is that either the scheduler doesn't care about thread affinity and evensched_getcpu()can't see that, OR, Oprofile falsely accounts samples to the wrong CPU. Or to the wrong symbol, although I've added "__attribute__((noinline))" to every called function to make sure that inlineing doesn't screw up anything.
I'm running my tests on Ubuntu Servers with 14.04 with 3.13.0-32 kernel, the CPU is a desktop kind, i5-4570 @ 3.20GHz (no-HT!) with oprofile 0.9.9. Anyone have a suggestion about what happens?Quoted 3 lines
Regards,
Zoltan
Hi Zoltan,
Are you using operf or opcontrol to set up the data measurements? It would be good to provide the command lines that you are using to do the data collection with oprofile.
"perf stat <command>" or "perf stat -p <
list_of_pids>" can show task migration. Look for the non-zero "cpu-migrations" in the output like in the output below.$ perf stat -p 29600 ^C Performance counter stats for process id '29600':
Quoted 14 lines
4303.761019 task-clock (msec) # 0.305 CPUs utilized [100.00%]
17,980 context-switches # 0.004 M/sec [100.00%]
595 cpu-migrations # 0.138 K/sec [100.00%]
44,212 page-faults # 0.010 M/sec
10,585,868,260 cycles # 2.460 GHz [100.00%]
6,463,554,435 stalled-cycles-frontend # 61.06% frontend cycles idle [100.00%]
<not supported> stalled-cycles-backend
10,400,529,789 instructions # 0.98 insns per cycle
# 0.62 stalled cycles per insn [100.00%]
2,049,813,299 branches # 476.284 M/sec [100.00%]
48,441,881 branch-misses # 2.36% of all branches
14.112614437 seconds time elapsedI've checked with perf, it doesn't show any cpu-migrations. I've also checked the actual vma addresses in the samples, they are indeed referring to that particular function. Tried to compile the application with -O0 and -Og, no luck so far
Quoted 3 lines
As a check you might try using "perf record <command>" or "perf record -p <
list_of_pids>". Once data collection is done use "perf report --sort=cpu" to see whether samples were recorded on other processors. That would determine how much the monitored process is beoing scheduled on other cpus.
I tried to use "perf report" in various ways, but I couldn't get any information out related to which CPU was used.
Quoted 7 lines
There is also a systemtap (https://sourceware.org/systemtap/) script
cycle_thief(https://sourceware.org/systemtap/examples/#process/cycle_thief.stp) that shows how often tasks are migrated and what other tasks are competing with the pid for time.These suggestions should help determine whether oprofile is getting the attribution of samples wrong or the scheduler is migrating the tasks.
-Will
Quoted 54 lines · Zoltan Kiss
Quoted 51 lines · William Cohen
Quoted 24 lines · Zoltan Kiss
Hi,
I'm using OProfile to check some suspicious behaviour of dpdk-pktgen, and I can see something which troubles me. Either the scheduler lies about core affinity or Oprofile accounts some samples wrongly. This userspace app runs in threads, which are assigned explicitly to one single core with
pthread_setaffinity_np(I've added some debug lines which checks the affinity withpthread_getaffinity_npto make sure that the affinity was correctly set indeed)
These threads run infinite loops, calling different functions. In my example one of these functions (pktgen_main_rxtx_loop) should only be called by the thread on core #1, but in the oprofile results I can see samples from all other cores as well. I've added asched_getcpu()call to every iteration of that infinite loop to check if it runs anywhere else than core #1, but it seems to be fine.
So my conclusion is that either the scheduler doesn't care about thread affinity and evensched_getcpu()can't see that, OR, Oprofile falsely accounts samples to the wrong CPU. Or to the wrong symbol, although I've added "__attribute__((noinline))" to every called function to make sure that inlineing doesn't screw up anything.
I'm running my tests on Ubuntu Servers with 14.04 with 3.13.0-32 kernel, the CPU is a desktop kind, i5-4570 @ 3.20GHz (no-HT!) with oprofile 0.9.9. Anyone have a suggestion about what happens?Quoted 3 lines
Regards,
Zoltan
Hi Zoltan,
Are you using operf or opcontrol to set up the data measurements? It would be good to provide the command lines that you are using to do the data collection with oprofile.
"perf stat <command>" or "perf stat -p <
list_of_pids>" can show task migration. Look for the non-zero "cpu-migrations" in the output like in the output below.$ perf stat -p 29600 ^C Performance counter stats for process id '29600':4303.761019 task-clock (msec) # 0.305 CPUs utilized [100.00%] 17,980 context-switches # 0.004 M/sec [100.00%] 595 cpu-migrations # 0.138 K/sec [100.00%] 44,212 page-faults # 0.010 M/sec 10,585,868,260 cycles # 2.460 GHz [100.00%] 6,463,554,435 stalled-cycles-frontend # 61.06% frontend cycles idle [100.00%] <not supported> stalled-cycles-backend 10,400,529,789 instructions # 0.98 insns per cycle # 0.62 stalled cycles per insn [100.00%] 2,049,813,299 branches # 476.284 M/sec [100.00%] 48,441,881 branch-misses # 2.36% of all branches
Quoted 6 lines
Quoted 2 lines
14.112614437 seconds time elapsedI've checked with perf, it doesn't show any cpu-migrations. I've also checked the actual vma addresses in the samples, they are indeed referring to that particular function. Tried to compile the application with -O0 and -Og, no luck so far
Hi Zoltan,
Could you write up the steps used to produce the problem so we have a reproducer and include the header and the portion of the opreport output that shows the problem?
-Will
Quoted 24 lines
Quoted 3 lines
As a check you might try using "perf record <command>" or "perf record -p <
list_of_pids>". Once data collection is done use "perf report --sort=cpu" to see whether samples were recorded on other processors. That would determine how much the monitored process is beoing scheduled on other cpus.I tried to use "perf report" in various ways, but I couldn't get any information out related to which CPU was used.
Quoted 7 lines
There is also a systemtap (https://sourceware.org/systemtap/) script
cycle_thief(https://sourceware.org/systemtap/examples/#process/cycle_thief.stp) that shows how often tasks are migrated and what other tasks are competing with the pid for time.These suggestions should help determine whether oprofile is getting the attribution of samples wrong or the scheduler is migrating the tasks.
-Will
------------------------------------------------------------------------------
New Year. New Location. New Benefits. New Data Center in Ashburn, VA. GigeNET is offering a free month of service with a new server in Ashburn. Choose from 2 high performing configs, both with 100TB of bandwidth. Higher redundancy.Lower latency.Increased capacity.Completely compliant.
http://p.sf.net/sfu/gigenet
_______________________________________________
oprofile-list mailing list [email protected]
https://lists.sourceforge.net/lists/listinfo/oprofile-list
Hi,
Another fun fact I forgot to mention: if I profile only this thread, and then run opreport with cpu filtering, it only shows results on the right cpu. So the wrong accounting only happens when I do system wide profiling. But not just with this thread and function, other functions appear at the wrong place as well.
Regards,
Zoltan
Quoted 63 lines · William Cohen
Quoted 24 lines · Zoltan Kiss
Hi,
I'm using OProfile to check some suspicious behaviour of dpdk-pktgen, and I can see something which troubles me. Either the scheduler lies about core affinity or Oprofile accounts some samples wrongly. This userspace app runs in threads, which are assigned explicitly to one single core with
pthread_setaffinity_np(I've added some debug lines which checks the affinity withpthread_getaffinity_npto make sure that the affinity was correctly set indeed)
These threads run infinite loops, calling different functions. In my example one of these functions (pktgen_main_rxtx_loop) should only be called by the thread on core #1, but in the oprofile results I can see samples from all other cores as well. I've added asched_getcpu()call to every iteration of that infinite loop to check if it runs anywhere else than core #1, but it seems to be fine.
So my conclusion is that either the scheduler doesn't care about thread affinity and evensched_getcpu()can't see that, OR, Oprofile falsely accounts samples to the wrong CPU. Or to the wrong symbol, although I've added "__attribute__((noinline))" to every called function to make sure that inlineing doesn't screw up anything.
I'm running my tests on Ubuntu Servers with 14.04 with 3.13.0-32 kernel, the CPU is a desktop kind, i5-4570 @ 3.20GHz (no-HT!) with oprofile 0.9.9. Anyone have a suggestion about what happens?Quoted 3 lines
Regards,
Zoltan
Hi Zoltan,
Are you using operf or opcontrol to set up the data measurements? It would be good to provide the command lines that you are using to do the data collection with oprofile.
"perf stat <command>" or "perf stat -p <
list_of_pids>" can show task migration. Look for the non-zero "cpu-migrations" in the output like in the output below.$ perf stat -p 29600 ^C Performance counter stats for process id '29600':4303.761019 task-clock (msec) # 0.305 CPUs utilized [100.00%] 17,980 context-switches # 0.004 M/sec [100.00%] 595 cpu-migrations # 0.138 K/sec [100.00%] 44,212 page-faults # 0.010 M/sec 10,585,868,260 cycles # 2.460 GHz [100.00%] 6,463,554,435 stalled-cycles-frontend # 61.06% frontend cycles idle [100.00%] <not supported> stalled-cycles-backend 10,400,529,789 instructions # 0.98 insns per cycle # 0.62 stalled cycles per insn [100.00%] 2,049,813,299 branches # 476.284 M/sec [100.00%] 48,441,881 branch-misses # 2.36% of all branches 14.112614437 seconds time elapsedAs a check you might try using "perf record <command>" or "perf record -p <
list_of_pids>". Once data collection is done use "perf report --sort=cpu" to see whether samples were recorded on other processors. That would determine how much the monitored process is beoing scheduled on other cpus.There is also a systemtap (https://sourceware.org/systemtap/) script
cycle_thief(https://sourceware.org/systemtap/examples/#process/cycle_thief.stp) that shows how often tasks are migrated and what other tasks are competing with the pid for time.These suggestions should help determine whether oprofile is getting the attribution of samples wrong or the scheduler is migrating the tasks.
-Will
Quoted 57 lines · William Cohen
Quoted 56 lines · Zoltan Kiss
Quoted 53 lines · William Cohen
Quoted 24 lines · Zoltan Kiss
Hi,
I'm using OProfile to check some suspicious behaviour of dpdk-pktgen, and I can see something which troubles me. Either the scheduler lies about core affinity or Oprofile accounts some samples wrongly. This userspace app runs in threads, which are assigned explicitly to one single core with
pthread_setaffinity_np(I've added some debug lines which checks the affinity withpthread_getaffinity_npto make sure that the affinity was correctly set indeed)
These threads run infinite loops, calling different functions. In my example one of these functions (pktgen_main_rxtx_loop) should only be called by the thread on core #1, but in the oprofile results I can see samples from all other cores as well. I've added asched_getcpu()call to every iteration of that infinite loop to check if it runs anywhere else than core #1, but it seems to be fine.
So my conclusion is that either the scheduler doesn't care about thread affinity and evensched_getcpu()can't see that, OR, Oprofile falsely accounts samples to the wrong CPU. Or to the wrong symbol, although I've added "__attribute__((noinline))" to every called function to make sure that inlineing doesn't screw up anything.
I'm running my tests on Ubuntu Servers with 14.04 with 3.13.0-32 kernel, the CPU is a desktop kind, i5-4570 @ 3.20GHz (no-HT!) with oprofile 0.9.9. Anyone have a suggestion about what happens?Quoted 3 lines
Regards,
Zoltan
Hi Zoltan,
Are you using operf or opcontrol to set up the data measurements? It would be good to provide the command lines that you are using to do the data collection with oprofile.
"perf stat <command>" or "perf stat -p <
list_of_pids>" can show task migration. Look for the non-zero "cpu-migrations" in the output like in the output below.$ perf stat -p 29600 ^C Performance counter stats for process id '29600':4303.761019 task-clock (msec) # 0.305 CPUs utilized [100.00%] 17,980 context-switches # 0.004 M/sec [100.00%] 595 cpu-migrations # 0.138 K/sec [100.00%] 44,212 page-faults # 0.010 M/sec 10,585,868,260 cycles # 2.460 GHz [100.00%] 6,463,554,435 stalled-cycles-frontend # 61.06% frontend cycles idle [100.00%] <not supported> stalled-cycles-backend 10,400,529,789 instructions # 0.98 insns per cycle # 0.62 stalled cycles per insn [100.00%] 2,049,813,299 branches # 476.284 M/sec [100.00%] 48,441,881 branch-misses # 2.36% of all branches 14.112614437 seconds time elapsed
Quoted 8 lines
Quoted 4 lines
I've checked with perf, it doesn't show any cpu-migrations. I've also checked the actual vma addresses in the samples, they are indeed referring to that particular function. Tried to compile the application with -O0 and -Og, no luck so far
Hi Zoltan,
Could you write up the steps used to produce the problem so we have a reproducer and include the header and the portion of the opreport output that shows the problem?
Hi,
Sure, but you need an Intel DPDK compatible NIC in your box:
Install DPDK:
cd /local
git clone http://dpdk.org/git/dpdk
git checkout v1.7.1
sed -i s/"CONFIG_RTE_BUILD_COMBINE_LIBS=n"/"CONFIG_RTE_BUILD_COMBINE_LIBS=y"/ config/common_linuxapp make config T=x86_64-native-linuxapp-gcc EXTRA_CFLAGS="-g"
make install T=x86_64-native-linuxapp-gcc EXTRA_CFLAGS="-g -fPIC -I/usr/include/x86_64-linux-gnu"
Configure the requirements:
sudo mkdir /mnt/huge
sudo mount -t hugetlbfs nodev /mnt/huge
sudo /sbin/modprobe uio
sudo insmod /local/dpdk/x86_64-native-linuxapp-gcc/kmod/igb_uio.ko ulimit -Sn 2048
echo 2048 > /sys/devices/system/node/node0/hugepages/hugepages-2048kB/nr_hugepages /local/dpdk/tools/dpdk_nic_bind.py --bind=igb_uio <ethdevice>
Install dpdk-pktgen:
git clone git://dpdk.org/apps/pktgen-dpdk
export RTE_SDK=/local/dpdk
export RTE_TARGET=x86_64-native-linuxapp-gcc
Start it like this:
app/build/pktgen -c 0xf -n 4 --proc-type auto -w <PCI id> -- -m "[1:2].0"
The following commands should be enough to start emitting packets:
set 0 size 64
proto udp 0
start 0
In the system-wide oprofile (see my commands in an another mail) you will see that e.g. pktgen_main_rx_loop can be see not just on CPU #1.
Regards,
Zoltan
Quoted 27 lines
-Will
Quoted 24 lines
Quoted 3 lines
As a check you might try using "perf record <command>" or "perf record -p <
list_of_pids>". Once data collection is done use "perf report --sort=cpu" to see whether samples were recorded on other processors. That would determine how much the monitored process is beoing scheduled on other cpus.I tried to use "perf report" in various ways, but I couldn't get any information out related to which CPU was used.
Quoted 7 lines
There is also a systemtap (https://sourceware.org/systemtap/) script
cycle_thief(https://sourceware.org/systemtap/examples/#process/cycle_thief.stp) that shows how often tasks are migrated and what other tasks are competing with the pid for time.These suggestions should help determine whether oprofile is getting the attribution of samples wrong or the scheduler is migrating the tasks.
-Will
------------------------------------------------------------------------------
New Year. New Location. New Benefits. New Data Center in Ashburn, VA. GigeNET is offering a free month of service with a new server in Ashburn. Choose from 2 high performing configs, both with 100TB of bandwidth. Higher redundancy.Lower latency.Increased capacity.Completely compliant.
http://p.sf.net/sfu/gigenet
_______________________________________________
oprofile-list mailing list [email protected]
https://lists.sourceforge.net/lists/listinfo/oprofile-list