Hacker News Reader

Comment by makmanalp | Hacker News Reader

This. It can be faster per-node because of removed IO and abstraction overheads, but perhaps not overall (and they allude to that fact). There are other side effects too though - you're limited by how large hardware on a single node can go, and what the cost/performance ratio is with highly specialized hardware like the above or commodity machines.

People also might want distributed systems for other reasons - fault tolerance is one. So as your SLA requirements go higher, you sometimes end up with little choice but to pay that cost. Other times you don't.

They do cite challenges with scaling this up, but I think quite a bit of their innovation is around just leveraging hardware better, which is a big enough deal on its own.

View on HN

In general, with current systems, it's all a partitioning problem. Speed-up is a trade-off between granularity of the compute kernel placed on a resource, the communication between said resources, latency at each level of the memory hierarchy (cache line utilization and reuse have a direct impact on this), and also (eek) protocol overheads between all the abstraction layers. Current systems are largely built around single CPU abstractions pasted on to make them multi-CPU abstractions. Working with accelerators, RDMA, etc. is quite a challenge both for the programmer manually making things work and for systems developers beating the abstractions into submission.

But can we have a cluster of computers that look like a single computer sharing the underlying hardware, because that's just basically linking a remote device. I supposed special kernel has to be written to overlay.

EDIT: Looks like this is already a thing [1]

[1]: https://en.wikipedia.org/wiki/Single_system_image

That's not quite the issue - SSI (which I didn't know about!) sounds cool, but it's a programming model that solves similar issues at a different abstraction layer. The question that remains unsolved is, regardless of what abstraction you use, whether the overhead of communication (latency, throughput) between nodes is large enough to make things annoying, and whether it's possible to work around this effectively or not. Stuff like InfiniBand along with tech like RDMA is helping a lot with this, and hardcore database installations do use this.

SSI is awesome, and (IIRC) something DragonflyBSD was working towards once upon a time.

Reply on news.ycombinator.com