Re: [hybi] frame length encoding

["Shelby Moore" <[e-mail]>]

> On Sat, Aug 21, 2010 at 4:27 PM, Pieter Hintjens <[e-mail]> wrote:
>
>> CPU core count increases but I think it's fair to say CPU speeds have
>> hit the top of the curve.  Thus we're constrained by the capacity of
>> one core to process the data off a network, while network capacity
>> continues to double every 18-24 months.

Why can't you utilize more than one core?

I realize we need a new computer language for multi-core and I started
working on that.  It is all about minimizing function side-effects, which
I call purity.  Copute is the name, but just in early idea stage.  I
realize such discussion is out-of-scope of this thread, and you may be
referring to real-time or other issues that I am not aware of.


> I'm not sure I would agree with you on the network bandwidth.  Many places
> I
> know of still use 100baseT to the desktop, 17 years after its introduction
> and are only slowly migrating to GigE (itself 11 years old) as
> infrastructure costs have come down to about equal.  I don't know of
> anyone
> running higher than GigE to the desktop.


My observations may not apply to more wealthy countries, but they probably
do apply to India and other developing countries in Latin America.

Last time I checked in 2009, afaiu that most of the low-end PC
motherboards (what sells here in Philippines in huge volumes), have buggy
wake-up on LAN if the GigeE is turned on, so many revert to the old
Ethernet.  I don't know if it is driver or OS issue.

In terms of mobile, here in Philippines where I am, which is perhaps #1
texting county in world, the 3G has very minimal coverage and in the large
population centers where there is coverage, it is overloaded and the
networks drop you back to EDGE/GPRS level of throughput after you hit
certain limits.  And during peak hours everything slows down here on every
consumer network.  There are no high end choices, except I think the call
centers somehow get the wide pipes they need.


>  Beyond that, anecdotally the
> typical size of the Internet connection at businesses doesn't seem to have
> increased since I was in the ISP business 10 years ago (home and mobile
> definitely has, though I still don't think at your suggested rate of
> doubling).
>
> I will grant your doubling pattern in Infiniband and Myrinet offerings,
> but
> I doubt any significant portion of WebSocket traffic will be running on
> such
> networks.
>
>
>> Indeed, there are probably very few people in the open source world
>> concerned with efficiency at the level we are.  I've explained our
>> experience, and will leave it at that.
>>
>
> It may also be that WebSocket, which is primarily designed to give
> TCP-like
> capabilities to web browsers, is not appropriate for your backend network
> --
> perhaps a frontend would speak WebSocket to end-users, and then a protocol
> more optimized for your needs going back to your servers.
>
> Anyway, I agree with you that getting agreement on the framing, as long as
> it is reasonable, is more important than what it actually is so we can get
> on to the next problem.