Hacker News Reader

Comment by viraptor | Hacker News Reader

Ok, I'm sceptical about the results. The reason is that there doesn't seem to be a massive difference between the tests. Since this fix is about speculative exec, why would it affect crypto code which is very register based and branchless as much as sqlite which is full of branches and memory/storage based? Why would it affect AES which is hardware accelerated as much as integer processing which is not?

I'm not saying this is impossible - maybe there's something that I'm missing. But it just doesn't add up at the moment. I'd love a more detailed / repeatable test.

View on HN

I can't speak about these benchmarks (the site is down). But Meltdown is not about speculative execution, it's about out-of-order execution. The fixes for Meltdown involve getting kernel data out of user virtual memory, and flushing the TLB if the processor does not have the pcid instruction. So context switches - including system calls - will tend to be more expensive, and if the processor does not implement pcid, way more expensive.

See https://meltdownattack.com/. I haven't yet read the Spectre paper, but I can say that the Meltdown paper is very readable and walks readers through everything. I anticipate the Spectre paper is as well.

This is about Spectre, not Meltdown, which is about speculative execution. From the article:

> Apple released iOS 11.2.2 update to address Spectre security issues.

I'm not sure why Meltdown was added to the HN title...

These benchmarks are between iOS 11.1.2 and 11.2.2, which include the battery throttling code in 11.2.0.

We need a benchmark between 11.2.1 and 11.2.2 instead.

Is it possible that this benchmark has coincided with a battery that's just crossing the threshold of battery-throttling induced slowdown?

The low scores in the article are in a similar ballpark to my (slow) iphone 6 on ios 10.3.3 with a 3yr old battery - i.e. about half what they should be with a fresh battery.

The A8 doesn't have a dedicated crypto engine like Intel chips do (uncore I think they call it). The A8 does have a couple of AES specific instructions so each core can speed up the work, but they were executed speculatively and presumably are no longer done so.

From Geekbench knowledgebase: "Geekbench will use AES instructions when available, and fall back to software implementations otherwise."

I expect this test to use them.

Reply on news.ycombinator.com