[dpdk-dev] KNI performance

[Jay Rolette]

The past few days I've been trying to chase down why operations over KNI
are so bloody slow. To give you an idea how bad it is, we did a simple test
over an NFS mount:

# Mount over a non-KNI interface (eth0 on vanilla Ubuntu 14.04 LTS)
$ time $(ls -last -R /mnt/sfs2008 > /dev/null)
real    11m58.224s
user    0m10.758s
sys     0m25.050s

# Reboot to make sure NFS cache is cleared and mount over a KNI interface
$ time $(ls -last -R /mnt/sfs2008 > /dev/null)
real    87m36.295s
user    0m14.552s
sys     0m25.949s

Packet captures showed a pretty consistent ~4ms delay. Get a READDIRPLUS
reply from NFS server and the TCP stack on the DPDK/KNI system took about
4ms to ACK the reply. It isn't just on ACK packets either. If there was no
ACK required, there would be a 4ms delay before the next call was sent
(ACCESS, LOOKUP, another READDIR, etc.).

This is running on top of a real DPDK app, so there are various queues and
ring-buffers in the path between KNI and the wire, so I started there. Long
story short, worst case, those could only inject ~120us of latency into the
path.

Next stop was KNI itself. Ignoring a few minor optos I found, nothing in
the code looked like it could account for 4ms of latency. That wasn't quite
right though...

Here's the code for the processing loop in kni_thread_single():

        while (!kthread_should_stop()) {
                down_read(&kni_list_lock);
                for (j = 0; j < KNI_RX_LOOP_NUM; j++) {
                        list_for_each_entry(dev, &kni_list_head, list) {
#ifdef RTE_KNI_VHOST
                                kni_chk_vhost_rx(dev);
#else
                                kni_net_rx(dev);
#endif
                                kni_net_poll_resp(dev);
                        }
                }
                up_read(&kni_list_lock);
                /* reschedule out for a while */
                schedule_timeout_interruptible(usecs_to_jiffies( \
                                KNI_KTHREAD_RESCHEDULE_INTERVAL));

Turns out the 4ms delay is due to the schedule_timeout() call. The code
specifies a 5us sleep, but the call only guarantees a sleep of *at least*
the time specified.

The resolution of the sleep is controlled by the timer interrupt rate. If
you are using a kernel from one of the usual Linux distros, HZ = 250 on
x86. That works out nicely to a 4ms period. The KNI kernel thread was going
to sleep and frequently not getting woken up for nearly 4ms.

We rebuilt the kernel with HZ = 1000 and things improved considerably:

# Mount over a KNI interface, HZ=1000
$ time $(ls -last -R /mnt/sfs2008 > /dev/null)

real    21m8.478s
user    0m13.824s
sys     0m18.113s

Still not where I'd like to get it, but much, much better.

Hopefully my pain is your gain and this helps other KNI users.

Jay