Comments on: AIO through stack scheduling http://www.zabbo.net/post/aio-through-stack-scheduling/ software princess Mon, 08 Sep 2008 12:30:41 +0000 http://wordpress.org/?v=2.3.1 By: zab http://www.zabbo.net/post/aio-through-stack-scheduling/#comment-14141 zab Tue, 12 Dec 2006 19:59:46 +0000 http://www.zabbo.net/post/aio-through-stack-scheduling/#comment-14141 > given how small is the path, it’s amazing. Well, it's just a prototype :). It is missing significant pieces. > 1) stack overflow (just 4K !!!) The current strategy copies the portion of the stack that a path is using into a buffer that is allocated at the time of the switch. It actually reduces stack consumption because data that is used by the waking path can't be on the sleeping stack. (That's a significant cost of the current approach.) > 2) having a stack per request Well, the part of the stack that's used when the path goes to sleep. I didn't notice what the dio read case was, but trivial stuff seems to be using hundreds of bytes. That can surely be trimmed down a bit -- the current stack switching function is absurdly conservative about saving registers. > given how small is the path, it’s amazing.

Well, it’s just a prototype :). It is missing significant pieces.

> 1) stack overflow (just 4K !!!)

The current strategy copies the portion of the stack that a path is using into a buffer that is allocated at the time of the switch. It actually reduces stack consumption because data that is used by the waking path can’t be on the sleeping stack. (That’s a significant cost of the current approach.)

> 2) having a stack per request

Well, the part of the stack that’s used when the path goes to sleep. I didn’t notice what the dio read case was, but trivial stuff seems to be using hundreds of bytes. That can surely be trimmed down a bit — the current stack switching function is absurdly conservative about saving registers.

]]> By: bzzz http://www.zabbo.net/post/aio-through-stack-scheduling/#comment-14140 bzzz Tue, 12 Dec 2006 19:53:18 +0000 http://www.zabbo.net/post/aio-through-stack-scheduling/#comment-14140 given how small is the path, it's amazing. my understanding is that "stack" in this context is a kernel task, literally. and it may result in: 1) stack overflow (just 4K !!!) 2) having a stack per request given how small is the path, it’s amazing.

my understanding is that “stack” in this context is a kernel task, literally. and it may result in:
1) stack overflow (just 4K !!!)
2) having a stack per request

]]> By: graydon http://www.zabbo.net/post/aio-through-stack-scheduling/#comment-14045 graydon Sat, 09 Dec 2006 01:37:12 +0000 http://www.zabbo.net/post/aio-through-stack-scheduling/#comment-14045 Very slick, nice work. Very slick, nice work.

]]>