[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] xen_evtchn_do_upcall

Hi Lan,

I am trying to improve the TCP/UDP performance in VM. Due the the vCPU scheduling on a pCPU sharing platform, the TCP receiving delay will be significant which hurt the TCP/UDP throughput. So I want to offload the softIRQ context to another Idle pCPU which I call fast-tick CPU. The packet receiving process is like this: IRQ routine can continue picking packet from ring buffer and put it to TCP receive buffer ( receive_queue, prequeue, backlog_queue ) in kernel no matter whether the user process on another CPU shared with other VMs is running or not. Once the vCPU holding the user process gets scheduled, user process will fetch all packets from receive buffer in kernel, which can improve the throughput. This works well for UDP unfortunately does not work for TCP currently.

I found those  xen_evtchn_do_upcall routines existing when irq context try to get the spinlock on the socket (Of course, they may happen in other paths). If this spinlock is held by process context, irq context has to spin on it and can not put any packet to receive buffer in time. So I doubt these   xen_evtchn_do_upcall routine are due to the synchronization between process context and irq context. Since I run process context and irq context on 2 different vCPU, when they try to get the spinlock on the same socket there will be interrupts between 2 vCPU which are implemented by event in Xen. 

If there is any error in my description, please correct me. Thanks.


2012/10/24 Ian Campbell <Ian.Campbell@xxxxxxxxxx>
On Wed, 2012-10-24 at 14:39 +0100, David Xu wrote:
> Hi Lan,
> Thanks for your reply. I did a experiment as follows,
> I assigned 2 vCPU to a VM, one vCPU (vCPU0) was pinned to a physical
> CPU shared with several other VMs and the other vCPU (vCPU1) was
> pinned to an idle physical CPU occupied by this VM only. Then in Guest
> OS I run iperf server to measure its TCP receiving throughput. In
> order to shrink the receiving delay caused by vCPU scheduling, I pin
> the IRQ context of NIC to vCPU1 and run iperf server on the vCPU0.
> This method works well for UDP, but does not work for TCP. I track the
> involved function by ftrace and get the following results which
> contains lots of xen_evtchn_do_upcall routine. What's the meaning of
> this process (xen_evtchn_do_upcall => handle_irq_event =>
> xennet_tx_buf_gc =>  gnttab_query_foreign_access)?

Have you looked at the code for any of those functions?

If you had done you'd find it is pretty obviously an interrupt being
delivered to the network device and the associated work to satisfy that

It doesn't seem that surprising that an iperf test should involve lots
of network interrupts.

It's not entirely clear to me what you are expecting to find and/or what
you are trying to prove.

> When I run both process context of application ( e.g. iperf server )
> and IRQ context on vCPU1 which is the ''fast" core, no
> any xen_evtchn_do_upcall routine found.

Perhaps on the fast core NAPI is able to kick in and therefore the NIC
becomes polled instead of interrupt driven?


Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.