Hacker News new | comments | show | ask | jobs | submit login

One of the approaches and the last Tyler, the speaker, introduced in the talk to improve the fat tail distribution of load balancing latency problem was Join-Idle-Queue algorithm[1] that came out of UIUC Extreme Computing Group, Microsoft Research and Azure in 2011.

And according to Tyler's data (fast.ly CDN data since he's the CTO there), it beats all the current load balancing algorithms including Join-Shortest-Queue. Six years later, I wonder:

a) was there any other novel / better algorithms tops JIQ in LB performance?

b) has Microsoft Azure LB uses JIQ internally in their LB offerings?

c) has any open source LB software implement JIQ algorithm on the horizon? (according to Tyler, he found none.)

Can someone with expertise share some lights on these? Thanks.

[1]: https://www.microsoft.com/en-us/research/wp-content/uploads/...




The JIQ algorithm seems to have very nice overlap with the work being done on reactive streams and reactive network protocols, for instance http://reactivesocket.io/

The idea being that, on top of a general purpose message stream, you overlay a back-channel where the stream receiver can signal to the producer how many more messages it's ready to handle.

Effectively what this does is moving the implicit back-pressure you get from network buffers up to an application level thing, where something like an app server can reason about how much load it's willing to take on.

Shameless plug: If you think that seems cool and are into Go, I wrote a client+server impl of ReactiveSocket that I'd love API feedback on from more experienced Go devs.. https://github.com/jakewins/reactivesocket-go


Google Scholar can come in handy here; there have been 117 papers citing JIQ and some of them may have improved on it: https://scholar.google.com/scholar?cites=3283470738712232937...


I am afraid that merely a citation reference doesn't tell the whole story. I guess someone has to dig into those papers and implement them and then use real world data to test and see, right?


If anyone does want to dig in to the papers, a friend and I made a discussion platform for research. I've added all the papers in Tyler's talk as well as two he recommended on Twitter to a list here: https://www.projectcredo.com/briankung/server-load-balancing

I have zero background in academic computer science, so it may take me a while to understand even one of these papers.


If a paper has something better than JIQ it should already contain benchmark results, so you could pick the paper with the best claimed performance and just implement that. The usual disclaimers about academic work apply.


I tested a ton of cdn networks. What i can say for sure is that the ms cdn is the fastest. So, i guess they use it for cdn at least.




Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | DMCA | Apply to YC | Contact

Search: