Hacker News new | ask | show | jobs
by bladerunner82 3748 days ago
Throughput. The BSD networking stack has been very carefully engineered over time. Linux is OK, but nowhere near as robust. FreeBSD is legendary for load tolerance and network throughput, that on equal machines would choke a Linux server. Seen it, work(ed) with it. When we do something serious that requires heavy load and awesome throughput, we always lean on FreeBSD, never anything else. Everything else has always let us down.
7 comments

I've seen this claim before, but I've also seen the counterclaim that Linux network stacks have since caught up and even surpassed the BSDs on throughput and latency.

Whether one or the other is truly better though might be irrelevant...either is such a monumental improvement over the windows networking stack that there are bound to be large benefits.

This is an OS for the control plane for hardware switches, so its throughput is not that important. Hardware does most of the heavy lifting.
I understand this, but was answering the question (in general) about why I believe BSD has better networking.
Not that I really doubt you (the FreeBSD developers have a hard won and worthy reputation in this area!) but do you have any benchmarks that help back this claim?

I promise this isn't a question to try to make a point, I'm genuinely interested.

Had mediocre network performance and increased load with freebsd on VMs (kvm), while at the same time a linux one would saturate a whole link.
Why are you trying to get network throughput on a VM? Of course it's going to be slow, but KVM probably has some passthrough optimisations for Linux guests.
Because that's what software defined networking is (mostly) about. In the majority of cases it's used as a way to setup a network plane for virtualized hosts.

Edit: See for example why VMWare bought Nicira: http://techcrunch.com/2012/07/23/vmware-buys-nicira-for-1-26...

I'm not a BSD guy, but I can still see it's a little unfair to compare the network throughputs of Linux and *BSD as guest VMs on a Linux host running KVM. You aren't going to get good throughout in that case without hardware passthrough (even then it's still bad). End of story.
I get your idea, but Kvm on linux is prolly the most deployed hypervisor in the world or will be soon. If it doesn't run well on it then it won't be used. End of story.
Sure it's unfair. But it is the current state of software.
That was as true ten years ago, is it still? Have there been any recent benchmarks?
I know that a bit more recently (~4 years ago) WhatsApp emphasized that FreeBSD's networking stack was part of what allowed them to scale so cheaply. I'm sure there was a good amount of benchmarking and profiling involved in that choice, using their production workloads.
Benchmarks please, thank you!
Got benchmark numbers on that?