Thanks for the overview Mike.  I just bailed up Gavin on IRC and between 
that convo and what you've just written I'm starting to picture a plan 
in my head... This sounds right up my alley, I wish I didn't have to go 
to bed right now as I've got a ton of ideas buzzing around I'd like to 
get started on right now.  But I'll be onto it as soon as I've got a 
free moment...

On 07/09/11 00:52, Mike Hearn wrote:
> On Tue, Sep 6, 2011 at 4:17 PM, Steve <shadders.del@gmail.com 
> <mailto:shadders.del@gmail.com>> wrote:
>
>     I'm not really understanding the use case though.  I believe most
>     bitcoind's have a default max connections of 8.  Is the goal to
>     increase this without fundamentally altering the bitcoind
>     concurrency model?
>
>
> bitcoind already uses asynchronous IO. That's not the problem.
>
> The issue came up in a conversation about scalability. If Bitcoins 
> popularity continues to grow, users are very likely to migrate away 
> from running full verifying nodes to lightweight clients, either a 
> different mode of the Satoshi client or different implementations like 
> the Android Wallet or MultiBit.
>
> Lightweight clients cannot verify thus should not relay. And they'll 
> be run by users who just want to send/receive coins from time to time, 
> so don't leave the programs running 24/7. The result could be running 
> out of sockets (like we have had problems with recently). It's 
> especially true because lightweight clients cannot check transactions 
> for themselves. If they want to show transactions appearing 
> immediately (and they do), they have to use "heard from lots of nodes" 
> as a proxy for validity. So lightweight clients are likely to be 
> socket intensive.
>
> We could solve this by just hoping that lots of people run full nodes. 
> The problem is that a full node is quite an intensive thing already, 
> it uses lots of CPU and disk seeks, and will just get more expensive 
> in future. And as transaction traffic increases, that leaves less CPU 
> time available to service thousands of connected clients. The ROI of 
> bringing up a new node decreases at the same time as the userbase 
> increases.
>
> One traditional approach to solving this is frontend proxies. 
> Jabber.com/org used this technique many years ago, and Google has also 
> used it to scale up the lockservice 
> <http://static.googleusercontent.com/external_content/untrusted_dlcp/labs.google.com/en/us/papers/chubby-osdi06.pdf> (see 
> section 3.1). It's effective because often maintaining connections to 
> thousands of clients doesn't involve much brainwork, just shifting 
> bytes around. This is especially true of Bitcoin. So if somebody is 
> running a full node already they could increase their client capacity 
> by just bringing up a frontend proxy and having it handle things like 
> outbound tx broadcasts/deduping inbound broadcasts, connection setup, 
> relaying recently found blocks etc. A well written proxy could 
> probably support tens of thousands of simultaneous clients which frees 
> up the bitcoinds time for verification and wallet manipulation.