On 16/02/17 20:08, Dumitrescu, Cristian wrote:
Hi Zoltan,
-----Original Message-----
From: dev [mailto:dev-boun...@dpdk.org] On Behalf Of Zoltan Kiss
Sent: Thursday, February 16, 2017 3:14 PM
To: dev@dpdk.org
Subject: [dpdk-dev] rte_sched library performance question
Hi,
I'm experimenting a little bit with the scheduler library, and I got some
performance numbers which seems to be worse than what I've expected.
I'm sending 64 bytes packets on a 10G interface to a separate thread, and
my simple test program (based on the qos_sched example) does the
following:
while (1) {
uint16_t ret = rte_ring_sc_dequeue_burst(it.ring,
(void**)flushbatch, FLUSH_SIZE);
rte_mbuf** t = flushbatch;
if (!ret) {
/* This call is necessary to make sure the TX completed
mbuf's
* are returned to the pool even if there is nothing to
* transmit */
rte_eth_tx_burst(it.portid, lcore, t, 0);
continue;
}
rte_sched_port_enqueue(it.port, flushbatch, ret);
ret = rte_sched_port_dequeue(it.port, flushbatch, FLUSH_SIZE);
Looks to me like the scheduler dequeue burst is equal to the enqueue burst size
of FLUSH_SIZE, right?
In this case, you are always dequeueuing the exact packets that you just
enqueued, and the scheduler dequeue needs to work really hard to find exactly
those FLUSH_SIZE queues that each one have a single packet at this point.
This is wht the enqueue burst size should be bigger than the dequeue burst
size. Basically, you add some water into the reservoir up to a reasonable fill
level before you start pouring it in your glass if you want to fill the glass
quickly.
Typical values used:
-for vector PMD: (enqueue = 32, dequeue = 24), (32, 28), (32, 16), etc
-for scalar PMD: (64, 48), (64, 32), ... We used (256, 248) for VPP
Thanks, it helped my case too. Btw. it would be good do link this
document somewhere in the DPDK docs, as it contains a lot of good
information about the scheduler:
https://networkbuilders.intel.com/docs/Network_Builders_RA_NFV_QoS_Aug2014.pdf
while (ret) {
uint16_t n = rte_eth_tx_burst(it.portid, lcore, t, ret);
/* we cannot drop the packets, so re-send */
/* update number of packets to be sent */
ret -= n;
t = &t[n];
};
}
I run this on a separate thread, another one doing rx and feeding the
packets to the ring. When I comment out the enqueue and dequeue part in
the
code (reducing it to simple l2fwd), I can forward the entire ~14 Mpps
traffic, whilst with the scheduler enabled I can only reach ~5.4 Mpps at
best. I've tried with a single pipe or with 4k (used rand() to randomly
distribute between pipe, everything else (class etc) was set to 0), didn't
make a difference. Is this expected? I'm running this on a Xeon E5-2630 0 @
2.30GHz
I've used the following configuration:
; port configuration [port]
[port]
frame overhead = 24
number of subports per port = 1
number of pipes per subport = 1024
queue sizes = 64 64 64 64
; Subport configuration
[subport 0]
tb rate = 1250000000; Bytes per second
tb size = 1000000000; Bytes
tc 0 rate = 1250000000; Bytes per second
tc 1 rate = 1250000000; Bytes per second
tc 2 rate = 1250000000; Bytes per second
tc 3 rate = 1250000000; Bytes per second
tc period = 10; Milliseconds
tc oversubscription period = 1000; Milliseconds
pipe 0-1024 = 0; These pipes are configured with pipe profile 0
; Pipe configuration
[pipe profile 0]
tb rate = 1250000000; Bytes per second
tb size = 1000000000; Bytes
tc 0 rate = 1250000000; Bytes per second
tc 1 rate = 1250000000; Bytes per second
tc 2 rate = 1250000000; Bytes per second
tc 3 rate = 1250000000; Bytes per second
tc period = 10; Milliseconds
tc 0 oversubscription weight = 1
tc 1 oversubscription weight = 1
tc 2 oversubscription weight = 1
tc 3 oversubscription weight = 1
tc 0 wrr weights = 1 1 1 1
tc 1 wrr weights = 1 1 1 1
tc 2 wrr weights = 1 1 1 1
tc 3 wrr weights = 1 1 1 1
Regards,
Zoltan
Regards,
Cristian
