tahoe-lafs/trac-2024-07-25

explore improved peer-selection approaches: chord, reliability-based #447

New issue

Open

opened 2008-06-03 05:12:56 +00:00 by warner · 0 comments

warner commented

2008-06-03 05:12:56 +00:00

Owner

Our old roadmap.txt had "Upload Peer Selection step 3/4" listed as "reliability/goodness-point counting?" and "denver airport (chord)?" listed.

What this means: our current "tahoe two" design could be changed to take historical server reliability into account. A long time ago, we imagined a system in which each server would get a certain number of points (based upon their demonstrated reliability, either computed by the uploading client itself, through other peers, or by a centralized service). Peer selection would assign shares to peers until it had reached a certain number of points. The idea was to take advantage of relatively unreliable peers instead of just always sticking with the reliable ones.

The "dernver airport / chord" idea involved reducing the number of connections needed to work with several files from fully-connected mesh down to log(N) (see also #235). We wrote this down somewhere. The rough idea was to send messages out along the chords towards the region near the SI, bearing a request to hold shares. The nodes which received and accepted the request would then connect directly to the uploading node to receive the share data.

Our old roadmap.txt had "Upload Peer Selection step 3/4" listed as "reliability/goodness-point counting?" and "denver airport (chord)?" listed. What this means: our current "tahoe two" design could be changed to take historical server reliability into account. A long time ago, we imagined a system in which each server would get a certain number of points (based upon their demonstrated reliability, either computed by the uploading client itself, through other peers, or by a centralized service). Peer selection would assign shares to peers until it had reached a certain number of points. The idea was to take advantage of relatively unreliable peers instead of just always sticking with the reliable ones. The "dernver airport / chord" idea involved reducing the number of connections needed to work with several files from fully-connected mesh down to log(N) (see also #235). We wrote this down somewhere. The rough idea was to send messages out along the chords towards the region near the SI, bearing a request to hold shares. The nodes which received and accepted the request would then connect directly to the uploading node to receive the share data.