On December 12, 2007, Robert Hailey wrote:
> 
> On Dec 11, 2007, at 5:59 PM, Matthew Toseland wrote:
> > Ok that could be interesting. Although ideally we'd have a
> > circular-keyspace-aware averager.
> >> ...
> > Ok. I suggest you commit, I will review post-commit.
> 
> Committed it r16508, though it has changed slightly according to my  
> experiments with swap-biasing (e.g. shows number of cache/store writes  
> as well).

Were should this appear?  Does it take a while to show up?  I have
r16509 running and am not seeing anything new (yet).

>  From what I have discovered (or theorized) thus far, using the  
> average location of the entire store to bias against is way too much  
> of an anchor. This is good, because it takes up way to much memory to  
> remember such a running average anyway. If we end up implementing such  
> a bias, in the end, it will likely just take into account the last few  
> inserts or successes (a small constant amount). In this way, the  
> pressure of your peers can still pull you into a new location  
> (dragging the weight of the last few inserts with you; which will be  
> updated).

Keeping the average key location and standard deviation just requires
a few counters along hooks to update them when adding or deleting the
store...  It also requires a single pass to get the initial numbers.  Should
not be expensive except for that first pass.

I agree that using the average would be very restrictive.  Average plus
or minus standard deviation would give a much better 'limit'.  I suspect
that It would take a fair length of time for it to start limiting at all (e.g. 
2 x standard deviation would be greater than 1).

> Whereas previously I have seen the 'storeDistance' to be 0.065 (IIRC),  
> after running with a recently-stored-location bias, I must have pulled  
> my peers closer to me as well, as now even with the patch off, I do  
> not see the storeDistance go nearly so high (staying around 0.004).  
> Or, since the network has no anchor, maybe I've pressured the whole  
> wheel to stop turning. For at least my node; this (valuing the store  
> as one peer) has not shown any increase in store hit rates, biasing  
> against the succeeding requests, however, has in the past raised it to  
> about 1%; but I would be even more concerning with implementing that.
> 
> With these stats I have noticed what may be an odd constant-location- 
> shift between what the node is requested to insert and retrieve, like  
> the network does not look for data in the exact same place it stores  
> it?!?!?! Suspect, but still can't confirm anything yet.

Ed


Reply via email to