I found out the cause of context-switches and iTLB misses. I ran the socket reading application on the same host with dpdk app. Sorry guys I was a fool :) Thank you for your help. Shawn you was right from the beginning I should've think more carefully about other applications running on the same host.
2016-04-15 1:47 GMT+03:00 ????????? ??????? <kiselev99 at gmail.com>: > Yes. 31% is ITLB-load-misses. My cpu is Intel(R) Core(TM) i5-2400 CPU @ > 3.10GHz. > There is no other big differences. I would say there is no other little > differences too. The numbers are about the same. > > I also noticed another strange thing: once I start sockets operations > perf context-switches counter increases in the ALL sibling thread > corresponded to dpdk lcores. But why? Only one thread is doing socket > operation and invoke system calls, so I expected context-switches to occur > only in that thread, not in the all threads. > > 2016-04-15 1:09 GMT+03:00 Hu, Xuekun <xuekun.hu at intel.com>: > >> I think 31.09% means ITLB-load-misses, right? To be strait forward, yes, >> this count means code misses is high that code footprint is big. For >> example, function A call function B, while the code address of B is far >> away from A. >> >> >> >> Is there any other big difference? Like L2/L3 cache miss? Actually I >> don?t expect the iTLB-load-misses could have that big impact (10% packet >> loss). >> >> >> >> BTW. What?s your CPU? >> >> >> >> *From:* ????????? ??????? [mailto:kiselev99 at gmail.com] >> *Sent:* Friday, April 15, 2016 5:33 AM >> *To:* Hu, Xuekun >> >> *Cc:* users at dpdk.org >> *Subject:* Re: [dpdk-users] Lcore impact >> >> >> >> I've done my homework with perf and the results show that >> iTLB-load-misses value is very high. In the tests without socket operations >> the processing lcore has 0.87% of all iTLB cache hits and there is no >> packet loss. In the test WITH socket operations the processing lcore >> has 31.09% of all iTLB cache hits and there is about 10% packet loss. How >> to interpret with results? Google shows a little about iTLB. So far some >> web pages suggest the following: >> >> "Try to minimize the size of the source code and locality so that >> instructions span a minimum number of pages, and so that the instruction >> span is less then the number of ITLB entries." >> >> >> >> Any ideas? >> >> >> >> >> >> >> >> >> >> >> >> 2016-04-14 23:43 GMT+03:00 Hu, Xuekun <xuekun.hu at intel.com>: >> >> Perf could. Or PCM, that is also a good tool. >> https://software.intel.com/en-us/articles/intel-performance-counter-monitor-a-better-way-to-measure-cpu-utilization >> >> >> >> *From:* ????????? ??????? [mailto:kiselev99 at gmail.com] >> *Sent:* Friday, April 15, 2016 3:31 AM >> *To:* Hu, Xuekun >> *Cc:* Shawn Lewis; users at dpdk.org >> >> >> *Subject:* Re: [dpdk-users] Lcore impact >> >> >> >> >> >> >> >> 2016-04-14 20:49 GMT+03:00 Hu, Xuekun <xuekun.hu at intel.com>: >> >> Are the two lcore belonging to one processor, or two processors? What the >> memory footprint is for the system call threads? If the memory footprint is >> big (>LLC cache size) and two locre are in the same processor, then it >> could have impact on packet processing thread. >> >> >> >> Those two lcores belong to one processor and it's a single processor >> machine. >> >> >> >> Both cores allocates a lot of memory and use the full dpdk arsenal: lpm, >> mempools, hashes and etc. But during the test the core doing socket data >> transfering is using only small 16k buffer for sending and sending is the >> all it does during the test. It doesn't use any other allocated memory >> structures. The processing core in turn is using rte_lpm whitch is big, but >> in my test there are only about 10 routes in it, so I think the amount >> "hot" memory is not very big. But I can't say if it's bigger than l3 cpu >> cache or not. Should I use some profilers and see if socket operations >> cause a lot of cache miss in the processing lcore? It there some tool that >> allows me to do that? perf maybe? >> >> >> >> >> >> >> >> -----Original Message----- >> From: users [mailto:users-bounces at dpdk.org] On Behalf Of Alexander >> Kiselev >> Sent: Friday, April 15, 2016 1:19 AM >> To: Shawn Lewis >> Cc: users at dpdk.org >> Subject: Re: [dpdk-users] Lcore impact >> >> I've already seen this documen and have used this tricks a lot of times. >> But this time I send data locally over localhost. There is even no nics >> bind to linux in my machine. Therefore there is no nics interruptions I can >> pin to cpu. So what do you propose? >> >> > 14 ???. 2016 ?., ? 20:06, Shawn Lewis <smlsr at tencara.com> ???????(?): >> > >> > You have to work with IRQBalancer as well >> > >> > >> http://www.intel.com/content/dam/doc/application-note/82575-82576-82598-82599-ethernet-controllers-interrupts-appl-note.pdf >> > >> > Is just an example document which discuss this (not so much DPDK >> related)... But the OS will attempt to balance the interrupts when you >> actually want to remove or pin them down... >> > >> >> On Thu, Apr 14, 2016 at 1:02 PM, Alexander Kiselev < >> kiselev99 at gmail.com> wrote: >> >> >> >> >> >>> 14 ???. 2016 ?., ? 19:35, Shawn Lewis <smlsr at tencara.com> ???????(?): >> >>> >> >>> Lots of things... >> >>> >> >>> One just because you have a process running on an lcore, does not >> mean thats all that runs on it. Unless you have told the kernel at boot to >> NOT use those specific cores, those cores will be used for many things OS >> related. >> >> >> >> Generally yes, but unless I start sending data to socket there is no >> packet loss. I did about 10 test runs in a raw and everythis was ok. And >> there is no other application running on that test machine that uses cpu >> cores. >> >> >> >> So the question is why this socket operations influence the other >> lcore? >> >> >> >>> >> >>> IRQBlance >> >>> System OS operations. >> >>> Other Applications. >> >>> >> >>> So by doing file i/o you are generating interrupts, where those >> interrupts get serviced is up to IRQBalancer. So could be any one of your >> cores. >> >> >> >> That is a good point. I can use cpu affinity feature to bind >> unterruption handler to the core not used in my test. But I send data >> locally over localhost. Is it possible to use cpu affinity in that case? >> >> >> >>> >> >>> >> >>> >> >>>> On Thu, Apr 14, 2016 at 12:31 PM, Alexander Kiselev < >> kiselev99 at gmail.com> wrote: >> >>>> Could someone give me any hints about what could cause permormance >> issues in a situation where one lcore doing a lot of linux system calls >> (read/write on socket) slow down the other lcore doing packet forwarding? >> In my test the forwarding lcore doesn't share any memory structures with >> the other lcore that sends test data to socket. Both lcores pins to >> different processors cores. So therotically they shouldn't have any impact >> on each other but they do, once one lcore starts sending data to socket the >> other lcore starts dropping packets. Why? >> >>> >> > >> >> >> >> >> >> -- >> >> ? ?????????, >> ??????? ????????? >> >> >> >> >> >> -- >> >> ? ?????????, >> ??????? ????????? >> > > > > -- > ? ?????????, > ??????? ????????? > -- ? ?????????, ??????? ?????????
