HI
In a run-to-completion simple model on a single core,
- you have full control of the execution order and packet
transmission order, typically being strictly the same than reception order.
- simple model to make it easy to debug and learn DPDK code.
- simple model to get deterministic latencies.
- will not scale easily, because the bottleneck is the whole sw.
In a simple pipeline model,
- execution order get less deterministic .
If the problems to solve are stateless (no dependency between
packets), that might be ok,
extreme scalability is easy to achieve.
- you need queues between pipelined components. You have to handle
overflows and bottlenecks,
and decisions to throttle and drop traffic is yours.
- you might wish to use priority queues, i.e. much more complex
design with its own drawbacks
and failure cases , and debugging uncertainties. Latencies get
hairy to analyse.
There are all intermediates between these extremes , e.g. pipelining
across multiple run-to-completion stateful subsystems
- stateful TCP reordering and reassembly,
- IP reordering and reassembly, IPSec , fragmentations between networks
with different mtus
- UDP reassembly before regex
- IGMP packets and multicast packets
- ...... all the networking fun .....
- and you have to design your subsystem to best fit workloads vs.
processing power of each core.
you could switch dynamically from one model to the other one depending
on traffic rates
- at low traffic rate, get the best use of a single core and save the
maximum of power on the system
this would favour a run-to-completion model at run time.
- at high traffic rate, wake up multiple cores
this would favour a piplining model at run time.
On 20/07/2020 16:08, Stephen Hemminger wrote:
On Mon, 20 Jul 2020 05:03:55 +0000
Prashanth Fernando <[email protected]> wrote:
Hi,
I'm wondering why DPDK proposes 2 different models.
Is there any situation that fits one model but not another?
I am looking to build an application with a firewall, regex, LPM, rate-limiters
etc ...
I am wondering which approach would be a best fit for my usecase.
Thanks,
Prashanth
There are two different factors here.
First, how many cores do you have to burn. The run to completion model uses
less cores (and has less latency). But other models are better if some set
of packets require longer to process (VPN, Crypto, ...) in that case you
want to push packets to other core.
--
Emutex Limited, Roselawn House, National Technology Park, Limerick, V94 6R68,
Ireland.
Phone: +353 (0)61 514496 Ext# 814, Web: www.emutex.com.
