From 011b7324366ec9b4c5546b42cb702b440ab7c672 Mon Sep 17 00:00:00 2001 From: Mike Perry Date: Thu, 25 Jun 2009 17:03:39 -0700 Subject: [PATCH] Update 161 to reflect current implementation. Also mention rounding step. --- .../161-computing-bandwidth-adjustments.txt | 72 ++++++++----------- 1 file changed, 31 insertions(+), 41 deletions(-) diff --git a/doc/spec/proposals/161-computing-bandwidth-adjustments.txt b/doc/spec/proposals/161-computing-bandwidth-adjustments.txt index b02dc64eb2..a50dba687e 100644 --- a/doc/spec/proposals/161-computing-bandwidth-adjustments.txt +++ b/doc/spec/proposals/161-computing-bandwidth-adjustments.txt @@ -42,31 +42,26 @@ Status: Open slices of 50 nodes each, grouped according to advertised node bandwidth. Two hop circuits are built using nodes from the same slice, and a large - file is downloaded via these circuits. For nodes in the first 15% of the - network, a 500K file will be used. For nodes in the next 15%, a 250K file - will be used. For nodes in next 15%, a 100K file will be used. The - remainder of the nodes will fetch a 75K file.[1] + file is downloaded via these circuits. The file sizes are set based + on node percentile rank as follows: + + 0-10: 4M + 10-20: 2M + 20-30: 1M + 30-50: 512k + 50-75: 256k + 75-100: 128k - This process is repeated 250 times, and average stream capacities are - assigned to each node from these results. - - In the future, a node generator type can be created to ensure that - each node is chosen to participate in an equal number of circuits, - and the selection will continue until every live node is chosen - to participate in at least 7 circuits. - + These sizes are based on measurements performed during test scans. -4. Ratio Calculation Options + This process is repeated until each node has been chosen to participate + in at least 5 circuits. - There are two options for deriving the ratios themselves. They can - be obtained by dividing each nodes' average stream capacity by - either the average for the slice, or the average for the network as a - whole. - Dividing by the network-wide average has the advantage that it will - account for issues related to unbalancing between higher vs lower - capacity, such as Steven Murdoch's queuing theory weighting result. - For this reason, we will opt for network-wide averages. +4. Ratio Calculation + + The ratios are calculated by dividing each measured value by the + network-wide average. 5. Ratio Filtering @@ -77,10 +72,8 @@ Status: Open with capacity of one standard deviation below a node's average are also removed. - The final ratio result will be calculated as the maximum of - these two resulting ratios if both are less than 1.0, the minimum - if both are greater than 1.0, and the mean if one is greater - and one is less than 1.0. + The final ratio result will be the unfiltered ratio if it is + close to 1.0, otherwise it will be the filtered ratio. 6. Pseudocode for Ratio Calculation Algorithm @@ -109,12 +102,7 @@ Status: Open Bw_net_ratio(N) = Bw_measured(N)/Bw_net_avg(Slices) Bw_Norm_net_ratio(N) = Bw_measured2(N)/Bw_Norm_net_avg(Slices) - if Bw_net_ratio(N) < 1.0 and Bw_Norm_net_ratio(N) < 1.0: - ResultRatio(N) = MAX(Bw_net_ratio(N), Bw_Norm_net_ratio(N)) - else if Bw_net_ratio(N) > 1.0 and Bw_Norm_net_ratio(N) > 1.0: - ResultRatio(N) = MIN(Bw_net_ratio(N), Bw_Norm_net_ratio(N)) - else: - ResultRatio(N) = MEAN(Bw_net_ratio(N), Bw_Norm_net_ratio(N)) + ResultRatio(N) = ClosestToOne(Bw_net_ratio(N), Bw_Norm_net_ratio(N)) 7. Security implications @@ -126,14 +114,14 @@ Status: Open This scheme will not address nodes that try to game the system by providing better service to scanners. The scanners can be detected - at the entry by IP address, and at the exit by the destination fetch. + at the entry by IP address, and at the exit by the destination fetch + IP. Measures can be taken to obfuscate and separate the scanners' source IP address from the directory authority IP address. For instance, scans can happen offsite and the results can be rsynced into the - authorities. The destination fetch can also be obscured by using SSL - and periodically changing the large document that is fetched. - + authorities. The destination server IP can also change. + Neither of these methods are foolproof, but such nodes can already lie about their bandwidth to attract more traffic, so this solution does not set us back any in that regard. @@ -148,7 +136,7 @@ Status: Open over a portion of the network, outputting files of the form: node_id= SP strm_bw= SP - filt_bw= NL + filt_bw= ns_bw= NL The most recent file from each scanner will be periodically gathered by another script that uses them to produce network-wide averages @@ -166,10 +154,15 @@ Status: Open scan, and taking the weighted average with the previous consensus bandwidth: - Bw_new = (Bw_current * Alpha + Bw_scan_avg*Bw_ratio)/(Alpha + 1) + Bw_new = Round((Bw_current * Alpha + Bw_scan_avg*Bw_ratio)/(Alpha + 1)) The Alpha parameter is a smoothing parameter intended to prevent - rapid oscillation between loaded and unloaded conditions. + rapid oscillation between loaded and unloaded conditions. It is + currently fixed at 0.333. + + The Round() step consists of rounding to the 3 most significant figures + in base10, and then rounding that result to the nearest 1000, with + a minimum value of 1000. This will produce a new bandwidth value that will be output into a file consisting of lines of the form: @@ -183,6 +176,3 @@ Status: Open This file can be either copied or rsynced into a directory readable by the directory authority. - -1. Exact values for each segment are still being determined via -test scans.