eRPC VS Seastar

The most common example is DPDK [1]. It's a framework for building bespoke networking stacks that are usable from userspace, without involving the kernel.

You'll find DPDK mentioned a lot in the networking/HPC/data center literature. An example of a backend framework that uses DPDK is the seastar framework [2]. Also, I recently stumbled upon a paper for efficient RPC networks in data centers [3].

If you want to learn more, the p99 conference by ScyllaDB has tons of speakers talking about some interesting challenges.

[1] https://www.dpdk.org/.

[2] https://github.com/scylladb/seastar

[3] https://github.com/erpc-io/eRPC

My side project has been to rewrite https://github.com/erpc-io/eRPC which does RPCs over UDP with some congestion control supposedly quite fast. Paper: https://www.usenix.org/system/files/nsdi19-kalia.pdf. I never got the code to work in AWS; I believe the author focused on Mellanox NICS but that's not really commodity H/W, which is where my interests lay.

So I dug into it .. and well I'll have my own library soon. I should be able to send UDP w/o congestion control sometime this week.

eRPC uses DPDK (100% user space NIC TX/RX control) plus the author's own other ideas to get performance. Since I'm getting into NICs + DPDK (in a serious way i.e this much more involved than vanilla sys/socket.h I/O) way, way late in the game, I hope and believe DPDK is, in the medium term, the better way to go than to turn to kernel improvements in for I/O.

Like others, getting the kernel out of the way, with pinned threads seems cleaner if one can develop from scratch.

This library will be a part of something bigger, however, a key architecture point for many people is: I got a RPC/packet/message. Now should I:

* process it in-place e.g. on the thread that was doing RX?

* delegate it to another core?

* if I delegate .. to whom?

* If I delegate how do I get a response back?

In DPDK I believe these are easier to decide and well-manage in code.

I used glommio as the async executor (instead of something like tokio), and it is wonderful. For people wondering whether it's "good enough" or to use C++ and seastar (as I have thought about a lot before starting this project), take the leap of faith, it's fast - both in terms of run time and to code.

b) It's proven now e.g Seastar, Glommio that the fastest way to run a multi-threaded application is to have one instance with one thread pinned per CPU core. Then to have fibers/lightweight threads on top handling all of the asynchronous code. Your approach of lots of instances is the slowest so there will be a ton of unnecessary thread context-switching.

The most common example is DPDK [1]. It's a framework for building bespoke networking stacks that are usable from userspace, without involving the kernel.

You'll find DPDK mentioned a lot in the networking/HPC/data center literature. An example of a backend framework that uses DPDK is the seastar framework [2]. Also, I recently stumbled upon a paper for efficient RPC networks in data centers [3].

If you want to learn more, the p99 conference by ScyllaDB has tons of speakers talking about some interesting challenges.

[1] https://www.dpdk.org/.

[2] https://github.com/scylladb/seastar

[3] https://github.com/erpc-io/eRPC

I assume Tokio itself, see e.g monoio or glommio, but also Seastar for C++.

https://core.dpdk.org/supported/ lists supported nics. You're best just reading material from the dpdk website for figuring out roughly what it is. It is used for a lot of different goals. For most web C++ stuff it's mainly used because you can avoid round trips of data passing through the kernel and can reference network data without tons of copying. For an example check out the SeaStar framework, https://seastar.io/, which is under the hood of ScyllaDB.

As this is a Kafka sub, this may be a good opportunity to mention that Redpanda is based on the same framework (seastar) as Scylla. The idea of sharding work to CPU cores turns out to apply very well to the Kafka data model, too!

Seastar which is a thread per core runtime written by the Scylla devs thats used in both Redpanda and Scylla as the underlying runtime. https://github.com/scylladb/seastar

ScyllaDB is, ironically, maybe one of the worst examples the author could have come up with for "abstraction" in the article.

If folks aren't familiar with their work/internal tech, go check out some of their repos like Seastar. They have some of the most talented systems programmers on the planet writing thin veneers over kernel and hardware API's to squeeze every ounce out of performance.

https://github.com/scylladb/seastar

I know it's beside the point, but I just had to share because I thought that was funny

I’ve seen frameworks for c++ (https://seastar.io/) and rust (https://github.com/actix/actix) which support what you’re describing out of the box.

If you're interested in scaling and asynchronous programming in c++ I highly recommend you investigate the SeaStar application framework. You wouldn't build a web service with SeaStar, rather you would build the infrastructure that you would use to build the web service on top of. https://github.com/scylladb/seastar