You just demonstrated one of the basic properties of vector packet processing:
as the offered load increases, the cost per vector element decreases. Although
you didn’t explicitly report the vector sizes involved, the vector size
necessarily increases as the offered load increases. Anyhow, it’s easy to fish
that that statistic out of the node runtime stats:
<start traffic>
“clear run”
<wait a while>
“show run”
<stop traffic>
You might ask: OK, why should the cost per packet decrease as the number of
packets in a vector increases? When you first enter a dispatch function, none
of the code involved is likely to be in the i-cache. The first packet incurs a
bunch of fixed overhead to drag code into the i-cache, and to warm up the
branch predictor. All of the other packets in the vector profit. On a
per-packet basis, cost decreases as the vector size increases.
There are a number of secondary effects with the same net result. Until the
vector size reaches 2, none of the graph nodes bother about prefetching. When
dealing with quad-looped nodes: ‘s/2/4/’.
This property gives rise to a second interesting property: given a specific
offered load and configuration, the vector size reaches a stable equilibrium.
Imagine the circuit time in equilibrium. Add a small delay [clock interrupt at
kernel level?] which increases the graph dispatch circuit time (rx ... process
... tx ... repeat).
The next rx vector size will be larger, but since it will be processed more
efficiently, the vector size will eventually return to the equilibrium value.
HTH... Dave
From: vpp-dev@lists.fd.io <vpp-dev@lists.fd.io> On Behalf Of Mikado
Sent: Thursday, November 22, 2018 10:22 PM
To: vpp-dev@lists.fd.io
Subject: [vpp-dev] Performance test issues : Average cpu time cost changes as
the tx speed changes.
Hi,
Recently I’m developing a plugin based on VPP 18.07 to decode specified
packets. I added it between dpdk-input and interface-output using the same
method of adding sample plugin in VPP source code. To test its performance in
theory , I used clib_cpu_time_now() to calculate the average cpu time cost
when packets go through my pulgin.When I use differnet tx speed to send
packets to the device using VPP and my plugin , it turns out that the average
cpu time changes as the tx speed changes. At frist , I assume it is caused by
the cpu time cost when VPP moves packets to the next node. So I calculate the
average cpu time of each node but it appears the same.Then, I use the sample
plugin to operate the same test. The result is similar although it does not
fluctuate much.
Now I’m confused. Isn’t it that all packets go through the same code and cost
the same cpu time ?
Here is my code and test result.
Code added in sample/node.c:
static uword
sample_node_fn (vlib_main_t * vm,
vlib_node_runtime_t * node,
vlib_frame_t * frame)
{
from = vlib_frame_vector_args (frame);
n_left_from = frame->n_vectors;
next_index = node->cached_next_index;
sample_main.last_cpu_time = clib_cpu_time_now ();
sample_main.total_pkts += n_left_from;
………
while (n_left_from > 0){
………
}
………
vlib_node_increment_counter (vm, sample_node.index,
SAMPLE_ERROR_SWAPPED, pkts_swapped);
sample_main.total_cpu_time += clib_cpu_time_now() - sample_main.last_cpu_time;
return frame->n_vectors;
}
Test result:
Tx speed(Mb/s) Average cpu time
190 14
475 12
665 9
950 7
1140 7
1900 6
2895 6
-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.
View/Reply Online (#11385): https://lists.fd.io/g/vpp-dev/message/11385
Mute This Topic: https://lists.fd.io/mt/28292345/21656
Group Owner: vpp-dev+ow...@lists.fd.io
Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com]
-=-=-=-=-=-=-=-=-=-=-=-
