aboutsummaryrefslogtreecommitdiff
path: root/src
diff options
context:
space:
mode:
authorPieter Wuille <pieter.wuille@gmail.com>2016-06-18 01:36:23 +0200
committerMatt Corallo <git@bluematt.me>2016-06-19 23:06:55 -0700
commitccd06b94f69c3e7758c35ac4bcd36d0e9450e158 (patch)
tree73158f8515144a523d87c00e6c797aac6ec589c3 /src
parent0d4cb48ef1b916679d9fad9f247a297c85c7fedf (diff)
Elaborate bucket size math
Diffstat (limited to 'src')
-rw-r--r--src/blockencodings.cpp14
1 files changed, 10 insertions, 4 deletions
diff --git a/src/blockencodings.cpp b/src/blockencodings.cpp
index 9a0805e400..7fd6a9cf52 100644
--- a/src/blockencodings.cpp
+++ b/src/blockencodings.cpp
@@ -85,10 +85,16 @@ ReadStatus PartiallyDownloadedBlock::InitData(const CBlockHeaderAndShortTxIDs& c
while (txn_available[i + index_offset])
index_offset++;
shorttxids[cmpctblock.shorttxids[i]] = i + index_offset;
- // Bucket selection is a simple Binomial distribution. If we assume blocks of
- // 10,000 transactions, allowing up to 12 elements per bucket should only fail
- // once every ~1.3 million blocks and once every 74,000 blocks in a worst-case
- // 16,000-transaction block.
+ // To determine the chance that the number of entries in a bucket exceeds N,
+ // we use the fact that the number of elements in a single bucket is
+ // binomially distributed (with n = the number of shorttxids S, and p =
+ // 1 / the number of buckets), that in the worst case the number of buckets is
+ // equal to S (due to std::unordered_map having a default load factor of 1.0),
+ // and that the chance for any bucket to exceed N elements is at most
+ // buckets * (the chance that any given bucket is above N elements).
+ // Thus: P(max_elements_per_bucket > N) <= S * (1 - cdf(binomial(n=S,p=1/S), N)).
+ // If we assume blocks of up to 16000, allowing 12 elements per bucket should
+ // only fail once per ~1 million block transfers (per peer and connection).
if (shorttxids.bucket_size(shorttxids.bucket(cmpctblock.shorttxids[i])) > 12)
return READ_STATUS_FAILED;
}