Subject: Re: Floating point in the kernel
To: Charles M. Hannum <email@example.com>
From: Jonathan Stone <jonathan@DSG.Stanford.EDU>
Date: 09/21/1998 19:14:21
>Now, can we talk about something that's actually relevant for a bit?
>There's enough buffering in the Winmodems that we just need to keep a
>low expected response time. This is really not much different in
>principal from driving a crappy PC serial chip, except that we have to
>do a bit more work on the CPU.
>`I don't see what the problem is.'
Recall that some people were talking about doing a sw modem in
userspace. that means the potential of pagefaults, context switches,
etc. That means additional buffering in userspace to keep the
in-kernel buffers filled.
That means more latency -- as much as the worst-case userspace
latency, e.g., a pagefault. Latency is Bad for a growing class of
applications, like Internet games or online multimedia or voice.
Modem packetization/compression latency is already a big hit even for