On 2019/11/13 2:50, William Tu wrote:
On Wed, Oct 30, 2019 at 5:32 PM Toshiaki Makita
<toshiaki.makita1@xxxxxxxxx> wrote:
On 2019/10/28 4:17, David Miller wrote:
From: Toke Høiland-Jørgensen <toke@xxxxxxxxxx>
Date: Sun, 27 Oct 2019 16:24:24 +0100
The results in the paper also shows somewhat disappointing performance
for the eBPF implementation, but that is not too surprising given that
it's implemented as a TC eBPF hook, not an XDP program. I seem to recall
that this was also one of the things puzzling to me back when this was
presented...
Also, no attempt was made to dyanamically optimize the data structures
and code generated in response to features actually used.
That's the big error.
The full OVS key is huge, OVS is really quite a monster.
But people don't use the entire key, nor do they use the totality of
the data paths.
So just doing a 1-to-1 translation of the OVS datapath into BPF makes
absolutely no sense whatsoever and it is guaranteed to have worse
performance.
1-to-1 translation has nothing to do with performance.
I think at least key size matters.
One big part of hot spots in xdp_flow bpf program is hash table lookup.
Especially hash calculation by jhash and key comparison are heavy.
The computational cost heavily depends on key size.
If umh can determine some keys won't be used in some way (not sure if it's
practical though), umh can load an XDP program which uses less sized
key. Also it can remove unnecessary key parser routines.
If it's possible, the performance will increase.
Toshiaki Makita
eBPF/XDP is faster only when you can by-pass/shortcut some code.
If the number of features required are the same, then an eBPF
implementation should be less than or equal to a kernel module's
performance. "less than" because eBPF usually has some limitations
so you have to redesign the data structure.
It's possible that after redesigning your data structure to eBPF,
it becomes faster. But there is no such case in my experience.
Regards,
William
