News for hackers. Stuff that matters.
Re:Hardware Store Elimination by bitwize on Wednesday May 13, 2020 @05:47PM (#23172213)
Isn't hardware store elimination when you order all your lumber, sheet metal, tools, etc. online and have them shipped straight to your door?
Re:Hardware Store Elimination by BeeOnRope on Wednesday May 13, 2020 @05:52PM (#23172265)
In retrospect, this was quite a poor name. I won't want anything to think their Home Depot will be taken away from them...

Perhaps someone can rename this to "Skylake CPU Store Elimination".

Re:Hardware Store Elimination by gumby on Wednesday May 13, 2020 @11:53PM (#23175077)
This addresses the case where you have he deliveries made directly to your shed, and you just get stuff from there.

For some reason he uses a French word for it though.

Re:Hardware Store Elimination by m463 on Wednesday May 13, 2020 @09:24PM (#23174091)
Judging from my projects, I would just trade 3 trips on saturday for 3 deliveries over a week. So I depend on small RTT.
Re:Hardware Store Elimination by shaklee3 on Thursday May 14, 2020 @12:54AM (#23175471)
Excellent article, as usual, Travis.
Re:Hardware Store Elimination by ylk1 on Wednesday May 13, 2020 @02:28PM (#23170043)
Nice read!
Re:Hardware Store Elimination by ylk1 on Wednesday May 13, 2020 @05:45PM (#23172198)
Can you try micro bench marking Apple Cores? Nobody seems to be doing that, yet people claim they are the best in business.
Re:Hardware Store Elimination by BeeOnRope on Wednesday May 13, 2020 @05:51PM (#23172254)
I don't have any Apple cores easily available, but the code [1] is open if anyone wants to try it (I don't know how POSIX-y the iOS compile environment is, though).

One caveat is that just because you don't find a performance difference, doesn't mean the optimization isn't happening. It could simply be the case that write throughput is not the limiter, but rather the latency * occupancy product is the limiter. E.g., if it takes 50 ns to go from L2 to RAM, and there are only 10 buffers available to hold these requests, then the maximum bandwidth is 64 bytes / 50 nanos * 10 buffers = 12.8 GB/s regardless of the maximum possible bandwidth of each component.

An eliminated store may still take this full latency (since it still has to read the value from RAM), so even if all writes are eliminated, the performance may remain at 12.8 GB/s - but you would save power and memory and L3 bandwidth for other cores... but the optimization would be tough to detect by looking at performance alone. You'd need to look at performance counters (does those exist for iOS devices?).

[1] https://github.com/travisdowns/zero-fill-bench