Re: Andrea Arcangeli's chat

Chemolli Francesco (USI) wrote:

> 1) O_DIRECT

Sounds nice. Should make quite a difference in efficiency of the OS
buffering if we disable buffering for swapouts.

For swapins we need the buffering, at least until Squid has a hot object
cache worth it's name..

> 2) Non-blocking, zero-copy sendfile()

This we are not very likely to be able to use in Squid. To use
sendfile() the disk and networking I/O must be tightly coupled, and we
are separating these two quite hard.. (intentional, clean design not
assuming how the store is built).

But it sounds very interesting for the people building high performance
servers of different kinds.. (web, ftp, file, ...). But I am a little
worried about implications in scalability and I/O bus bugs (PCI device
to device DMA is not commonly used, so many hardware bugs may be
expected there...)

> 3) sendto(MSG_MORE)
> should be a packetized high-efficiency interface to TCP stuff out.

Will probably come in handy when the store/protocol->client side
interface have been cleaned out, to quickly stuff out the headers and
first block of data.

--
Henrik