Re: [dpdk-users] Peformance troubleshouting of TCP/IP stack over DPDK.

[Stephen Hemminger <stephen@networkplumber.org>]

On Thu, 7 May 2020 07:09:44 -0700
dave seddon <dave.seddon.ca@gmail.com> wrote:

> tc qdisc
> https://linux.die.net/man/8/tc
> 
> On Thu, May 7, 2020 at 3:47 AM Pavel Vajarov <freakpv@gmail.com> wrote:
> 
> > On Wed, May 6, 2020 at 5:55 PM Stephen Hemminger <  
> > stephen@networkplumber.org>  
> > wrote:
> >  
> > > On Wed, 6 May 2020 08:14:20 +0300
> > > Pavel Vajarov <freakpv@gmail.com> wrote:
> > >  
> > > > Hi there,
> > > >
> > > > We are trying to compare the performance of DPDK+FreeBSD networking  
> > stack  
> > > > vs standard Linux kernel and we have problems finding out why the  
> > former  
> > > is  
> > > > slower. The details are below.
> > > >
> > > > There is a project called F-Stack <https://github.com/F-Stack/f-stack  
> > >.  
> > > > It glues the networking stack from
> > > > FreeBSD 11.01 over DPDK. We made a setup to test the performance of
> > > > transparent
> > > > TCP proxy based on F-Stack and another one running on Standard Linux
> > > > kernel.
> > > > We did the tests on KVM with 2 cores (Intel(R) Xeon(R) Gold 6139 CPU @
> > > > 2.30GHz)
> > > > and 32GB RAM. 10Gbs NIC was attached in passthrough mode.
> > > > The application level code, the one which handles epoll notifications  
> > and  
> > > > memcpy data between the sockets, of the both proxy applications is 100%  
> > > the  
> > > > same. Both proxy applications are single threaded and in all tests we
> > > > pinned the applications on core 1. The interrupts from the network card
> > > > were pinned to the same core 1 for the test with the standard Linux
> > > > application.
> > > >
> > > > Here are the test results:
> > > > 1. The Linux based proxy was able to handle about 1.7-1.8 Gbps before  
> > it  
> > > > started to throttle the traffic. No visible CPU usage was observed on  
> > > core  
> > > > 0 during the tests, only core 1, where the application and the IRQs  
> > were  
> > > > pinned, took the load.
> > > > 2. The DPDK+FreeBSD proxy was able to thandle 700-800 Mbps before it
> > > > started to throttle the traffic. No visible CPU usage was observed on  
> > > core  
> > > > 0 during the tests only core 1, where the application was pinned, took  
> > > the  
> > > > load. In some of the latter tests I did some changes to the number of  
> > > read  
> > > > packets in one call from the network card and the number of handled  
> > > events  
> > > > in one call to epoll. With these changes I was able to increase the
> > > > throughput
> > > > to 900-1000 Mbps but couldn't increase it more.
> > > > 3. We did another test with the DPDK+FreeBSD proxy just to give us some
> > > > more info about the problem. We disabled the TCP proxy functionality  
> > and  
> > > > let the packets be simply ip forwarded by the FreeBSD stack. In this  
> > test  
> > > > we reached up to 5Gbps without being able to throttle the traffic. We  
> > > just  
> > > > don't have more traffic to redirect there at the moment. So the  
> > bottlneck  
> > > > seem to be either in the upper level of the network stack or in the
> > > > application
> > > > code.
> > > >
> > > > There is a huawei switch which redirects the traffic to this server. It
> > > > regularly
> > > > sends arping and if the server doesn't respond it stops the  
> > redirection.  
> > > > So we assumed that when the redirection stops it's because the server
> > > > throttles the traffic and drops packets and can't respond to the arping
> > > > because
> > > > of the packets drop.
> > > >
> > > > The whole application can be very roughly represented in the following  
> > > way:  
> > > >  - Write pending outgoing packets to the network card
> > > > - Read incoming packets from the network card
> > > >  - Push the incoming packets to the FreeBSD stack
> > > >  - Call epoll_wait/kevent without waiting
> > > >  - Handle the events
> > > >  - loop from the beginning
> > > > According to the performance profiling that we did, aside from packet
> > > > processing,
> > > >  about 25-30% of the application time seems to be spent in the
> > > > epoll_wait/kevent
> > > > even though the `timeout` parameter of this call is set to 0 i.e.
> > > > it shouldn't block waiting for events if there is none.
> > > >
> > > > I can give you much more details and code for everything, if needed.
> > > >
> > > > My questions are:
> > > > 1. Does somebody have observations or educated guesses about what  
> > amount  
> > > of  
> > > > traffic should I expect the DPDK + FreeBSD stack + kevent to process in  
> > > the  
> > > > above
> > > > scenario? Are the numbers low or expected?
> > > > We've expected to see better performance than the standard Linux kernel  
> > > one  
> > > > but
> > > > so far we can't get this performance.
> > > > 2. Do you think the diffrence comes because of the time spending  
> > handling  
> > > > packets
> > > > and handling epoll in both of the tests? What do I mean. For the  
> > standard  
> > > > Linux tests
> > > > the interrupts handling has higher priority than the epoll handling and
> > > > thus the application
> > > > can spend much more time handling packets and processing them in the  
> > > kernel  
> > > > than
> > > > handling epoll events in the user space. For the DPDK+FreeBSD case the  
> > > time  
> > > > for
> > > > handling packets and the time for processing epolls is kind of equal. I
> > > > think, that this was
> > > > the reason why we were able to get more performance increasing the  
> > number  
> > > > of read
> > > > packets at one go and decreasing the epoll events. However, we couldn't
> > > > increase the
> > > > throughput enough with these tweaks.
> > > > 3. Can you suggest something else that we can test/measure/profile to  
> > get  
> > > > better idea
> > > > what exactly is happening here and to improve the performance more?
> > > >
> > > > Any help is appreciated!
> > > >
> > > > Thanks in advance,
> > > > Pavel.  
> > >
> > > First off, if you are testing on KVM, are you using PCI pass thru or  
> > SR-IOV  
> > > to make the device available to the guest directly. The default mode uses
> > > a Linux bridge, and this results in multiple copies and context switches.
> > > You end up testing Linux bridge and virtio performance, not TCP.
> > >
> > > To get full speed with TCP and most software stacks you need TCP
> > > segmentation
> > > offload.
> > >
> > > Also software queue discipline, kernel version, and TCP congestion  
> > control  
> > > can have a big role in your result.
> > >  
> >
> > Hi,
> >
> > Thanks for the response.
> >
> > We did the tests on Ubuntu 18.04.4 LTS (GNU/Linux 4.15.0-96-generic
> > x86_64).
> > The NIC was given to the guest using SR-IOV.
> > The TCP segmentation offload was enabled for both tests (standard Linux and
> > DPDK+FreeBSD).
> > The congestion control algorithm for both tests was 'cubic'.
> >
> > What do you mean by 'software queue discipline'?

The default qdisc in Ubuntu should be fq_codel (see tc qdisc show)
and that in general has a positive effect on reducing bufferbloat.

F-stack probably doesn't use TSO, you might want to look at TCP stack
from FD.io for comparison.