Yup, this is what you can do when you code directly against kernel bypass APIs like DDPK.
Important to note they also use effectively a shared nothing architecture between cores. This makes it very NUMA friendly and further favours network performance because you are able to lock processes/threads to the cores that are directly attached to said network adapter.
The architecture is pretty well known but it's nice to see someone build a framework out of it.
This. There are plenty of opportunities to try out userland/"kernel bypass" network stacks, but I've always wanted to see a framework that integrates it for me. Even if you still need to get elbow deep to do anything advanced, it's great to see a template, a skeleton, on how to put everything together.
They mention that this is using a polling architecture so each core is using 100% cpu even when idle. I've worked on high performance systems architected around polling before, but assumed that would be prohibitively costly these days now that cpu's have really advanced power management capabilities. Is polling architecture still viable for large scale cloud computing given costs of energy consumption? Years ago cpu's used a flat amount of power regardless of cpu utilization but now those extra cycles cost watts and of course money.
Perhaps such a server could be integrated in a way such that it spins up during high load scenarios and back down when the load is no longer sufficient to keep it saturated? If that were possible, then overall efficiency could still be higher than using only servers with better idle power usage characteristics since it could potentially save you from having to deploy n times as many normal servers.
Sounds awesome but I wouldn't look at something like this (which is awesome btw) unless I was expecting it be under massive sustained load.
For everything else I'd just use Python or PHP as they are mostly fast enough.
I am planning to write a framework in Rust when I have the time though as I find it an interesting language and as a web dev it's my area, I'll definitely be looking at how you approached stuff ;).
Cool, it would be an interesting design to see something with polling and perhaps different gears for lower load levels. BTW, the system with polling architecture that I mentioned I worked on is still in production and selling well(I hear).
Hmmm... wonder if this could go even faster using Cluster-on-Die mode in the new E5-2600v3 CPUs (It splits each CPU into 2 NUMA nodes).
Its also worth noting that the fastest memory configuration is 4x16GB DDR4-2133 DIMMs. Gets you the lowest latency and highest bandwidth according to Intel.
Another note, I see the CPU in the writeup reports as "cpu MHz: 1199.953" - have you turned off C-States?
Thanks. We "disable" C-states by polling -- the cpu never goes idle. This is partly because dpdk doesn't support interrupts, and partly because using interrupts is a major performance hit. We have plans to improve this, but nothing committed yet.
A lot of their speed comes from special tricks like bypassing the kernel and going direct to hardware (DPDK).
Intel's DPDK doesn't support Mac OS X.
You should probably develop in a VM anyways, or at least with a platform + hardware that supports these features so you can better match what production deployments will be like.
I always cringe a little, for backend systems software, when people develop in an environment completely divorced from their production environment. It just creates headaches.
In posix mode, seastar uses the host's TCP stack (instead of the internal super-fast stack). This is useful for development. While we haven't ported it to osx, it should be fairly easy as the code is portable.
Important to note they also use effectively a shared nothing architecture between cores. This makes it very NUMA friendly and further favours network performance because you are able to lock processes/threads to the cores that are directly attached to said network adapter.
The architecture is pretty well known but it's nice to see someone build a framework out of it.