Hacker News Reader

Comment by trendia | Hacker News Reader

Not to diminish this feat, but I wonder how the Phoronix benchmarks compare between an ARM cluster and a single x86 chip. Consider the OpenSSL benchmark.

A Sopine A64 compute module costs $29 and performs 15.05 OpenSSL signs / second [0].

An AMD Ryzen 1700 costs ~$300 and performs 986.73 OpenSSL signs / second [1].

So, if you spent $300 on 10 Pine64's, you would achieve approximately 150 signs / second, but for the same cost you could just buy the AMD chip and achieve approximately 6.5x better performance for the dollar.

[0] https://www.phoronix.com/scan.php?page=article&item=pine-64-...

[1] https://www.phoronix.com/scan.php?page=article&item=amd-ryze...

View on HN

I don't think openssl in phronix benchmarks use armv8 instructions, which makes those benchmarks unfair. There are benchmarks on cloudfare with openssl-devel, with hardware armv8, where arm performed decently. We have to wait and see how this pine64 does.

https://blog.cloudflare.com/arm-takes-wing/

armv7 and armv8 is just 64 bit, so I guess a corner case like openssl signs could benefit massively (lots of bigint) but normal workloads only marginally, while taking a massive hit on memory (don't want the 1 GiB model).

The ARM they tested at Cloudflare is also of course nothing like the mobile chip that powers the pine.

AArch64 is actually a completely new and much improved instruction set. More registers, less warts that inhibit out-of-order execution, etc.

This is entirely to lacking an optimized version for the new aarch64 instruction set. Similar problem noted in Go benchmarks awhile back, see 100x speed up here: https://blog.minio.io/accelerating-sha256-by-100x-in-golang-...

I suspect this is a more inexpensive/interesting way to get a small cluster together. (You could also look at user/surplus PCs, but then you probably will use a lot of space and power, and not necessary get a cluster of exactly the same boards and CPUs and what not)

This was my thought in looking at this post. I could get a several node cluster going with low power and small space and test various ideas out.

> I wonder how the Phoronix benchmarks compare between an ARM cluster and a single x86 chip.

I don't think the focus here would be performance, but the cluster operations. Also, having slow hardware helps with spotting some performance issues (there are things we get away with Xeons that we simply can't with ARMs) even though fast hardware will help uncover some concurrency bugs that wouldn't happen on ARM.

As for virtualizing a cluster inside a bigger Xeon box, I worked with server provisioning software and the kind of nasty bugs you find with real hardware is pretty surprising. Some nodes refuse to shut down, some refuse to start, some take longer than expected (please see http://bazaar.launchpad.net/~maas-committers/maas/trunk/revi... for the smallest commit ever - a single bit flip - that actually solves a bug) and so on.

I'll probably put mine inside an empty HP workstation pizza box I have with a proper power supply, along with a couple Pi's, a PiHAT cluster, a CHIO and an early Pine A64. Having a heterogeneous cluster is a necessity for testing.

Reply on news.ycombinator.com