[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <b6fcb866-f24e-4508-97f5-eb2b63f6eec7@gmail.com>
Date: Mon, 8 Jul 2024 10:45:46 +0200
From: Leone Fernando <leone4fernando@...il.com>
To: Eric Dumazet <edumazet@...gle.com>
Cc: davem@...emloft.net, kuba@...nel.org, pabeni@...hat.com,
dsahern@...nel.org, willemb@...gle.com, netdev@...r.kernel.org,
linux-kernel@...r.kernel.org
Subject: Re: [PATCH net-next v2 2/4] net: dst_cache: add input_dst_cache API
Hi Eric
> It targets IPv4 only, which some of us no longer use.
The same optimization can be implemented and applied to IPv6 as well and
I'm planning to do so.
> Also, what is the behavior (loss of performance) of this cache under
> flood, when 16 slots need to be looked up for each packet ?
I did some more measurements under flood as you asked, those are the results:
Total PPS:
number of IPs(x# cache size) mainline patched delta
Kpps Kpps %
20000(x40) 6498 6123 -5.7
35000(x70) 6502 5261 -19
50000(x100) 6503 4986 -23.3
I found out that most of the negative effect under extreme flood is
as a result of the atomic operations.
I have some ideas on how to solve this problem. it will demand a bit more memory
percpu, if that's reasonable I can do V3.
IMO the boost in the average case is worth the flood case penalty.
> This patch adds 10MB of memory per netns on host with 512 cpus,
> and adds netns creation and dismantle costs (45% on a host with 256 cpus)
Is such a memory penalty a deal breaker considering the performance gain?
Also I think adding a sysctl to control the size of the cache might be a good
solution for those who want to keep their memory usage lower. what do you think?
Thanks,
Leone
Powered by blists - more mailing lists