[Tech] Re: Re: The effect of slow nodes

[Jano]

toad wrote:

> On Fri, Dec 01, 2006 at 07:32:52PM +0100, Jano wrote:
>> toad wrote:
>> 
>> > Interesting. I'm not entirely sure what these graphs tell us though, as
>> > explained earlier ...
>> > 
>> > It is very surprising that slow nodes don't drag down the whole
>> > network;
>> 
>> Actually throughput seems lower and wilder; around 30k instead of 40k.
>> More runs could clarify this.
> 
> MRogers said there was a lot of variation between runs... more runs are
> therefore a good thing in any case.

I'll leave some running overnight.

>> 
>> > throttling without backoff seems competitive with, if not better than,
>> > throttling with backoff, even on a heterogenous network.
>> > 
>> > Ideas? Explanations?
>> 
>> I don't know what throttling and backoff are doing, so I'm unable to help
>> here.
> 
> Throttling = request senders slow down when they see overload messages
> or timeouts. Derived from TCP.

Slow down for all peers or only overloaded ones?

> Backoff = node doesn't route to another node for a while after a timeout
> or overload message. This is doubled the following time, until we reach
> a limit; it is reset if a request completes without overload or timeout.
> Derived from ethernet.

Thanks.

> The theory goes that if there are many slow nodes on the network,
> throttling would pull the whole network down to match their speed. This
> is why we have backoff; to prevent this from happening.

Some more data apart from success/failure must be extracted from these sims
to see what's really happening, I guess...