michaelkirkland.org/blog


On lots and lots of cores

Ars Technica reports on an Intel blog warning developers that we need to adapt an open ended number of cores.

Intel, of course, is primarily worried about making sure people are buying the n-core chips they'll be selling in the years ahead. Of course, that doesn't mean they're wrong, but I don't think the changes, from a coder's view, are going to be as generalized as some seem to. You're not going to get the people smearing their VB on the walls or poking Sharepoint with a stick to wrangle threads. Most of those folks can't even handle pointers without cutting themselves. If they're to see any benefit, it'll have to be done for them at a lower level, and that's fine.

Now, I'm not saying we won't see big changes in how we code. We certainly will. My point in this post is that it doesn't matter. We're going to go through interesting times, and there will be lots of attempts at getting parallelization right, but this is a revolutionary rather than evolutionary change.

The really cool stuff will spring off from the side, where no one was looking. Ars correctly points out that we won't be getting "free" performance upgrades in terms of periodic increases in clock speed anymore. What's important to note is that we will, suddenly, start getting "free" processors no one really cares about because they're idling.

Expect filesystems to get a lot smarter. Need to clear IO cache? Throw a spare core at compressing it rather than just tossing it. This is easy to parallelize, so throw all the idle cores at it.

Expect virtualization to get thrown at all sorts of problems. Need backwards compatibility? Keep whatever you need running on a core in the background.

You'll likely only be running local servers for sensitive or frequently accessed large stores of data. Renting virtual server instances is going to get cheap. When you can fit a few hundred cores into 1U, the price of renting one will probably be rolled into the cost of bandwidth.

Keep in mind, Intel is far from wrong. We still need to find ways to sensibly use lots of cores for singular tasks, but the really neat things will come out of the slack that arises when we don't keep up with them.

Bookmark and Share

Caunteecage said on 2013-05-07 17:55:04

http://www.maps.google.com/ - http://www.wikipedia.org/ - wiki
mail


yooying said on 2018-01-10 07:33:02

Thank you for sharing. Everything changes day by day, and coder has to change also.


net worth said on 2018-04-27 11:57:51

Nothing could last forever, especially Intel chips.