Comparing the performance of distributed hash tables under churn
Jinyang Li, Jeremy Stribling, Thomer M. Gil, Robert Morris, M. Frans Kaashoek
MIT Computer Science and Artificial Intelligence Laboratory
jinyang, strib, thomer, rtm, kaashoek
A protocol for a distributed hash table (DHT) incurs communi-
cation costs to keep up with churn—changes in membership—in
order to maintain its ability to route lookups efficiently. This pa-
per formulates a unified framework for evaluating cost and per-
formance. Communication costs are combined into a single cost
measure (bytes), and performance benefits are reduced to a single
latency measure. This approach correctly accounts for background
maintenance traffic and timeouts during lookup due to stale routing
data, and also correctly leaves open the possibility of different pref-
erences in the tradeoff of lookup time versus communication cost.
Using the unified framework, this paper analyzes the effects of DHT
parameters on the performance of four protocols under churn.
The design space of DHT protocols is large. While all designs
are similar in that nodes forward lookups for keys through
routing tables that point to other nodes, algorithms differ in
the amount of state they keep: from
with respect to a
network size of size
[7, 9] to
[10, 13, 14, 16] to
. They also differ in the techniques
used to find low latency routes, in the way they find alternate
paths after encountering dead intermediate nodes, in the ex-
pected number of hops per lookup, and in choice of parame-
ters such as the frequency with which they check other nodes
How is one to compare these protocols in a way that sep-
arates incidental details from more fundamental differences?
Most evaluations and comparisons of DHTs have focused on
lookup hopcount latency, or routing table size in unchanging
networks [2, 12, 15]. Static analysis, however, may unfairly
favor protocols that keep large amounts of state, since they
pay no penalty to keep the state up to date, and more state
usually results in lower lookup hopcounts and latencies.
This paper presents a framework for evaluating DHT al-
gorithms in the face of joining and leaving nodes, in a way
that makes it easy to compare tradeoffs between state main-
tenance costs and lookup performance. The paper compares
the Tapestry , Chord , Kelips , and Kademlia 
), supported by the National Sci-
ence Foundation under Cooperative Agreement No. ANI-0225660.
lookup algorithms within this framework. These four reflect
a wide range of design choices for DHTs.
We have implemented a simple simulator that models