1. Simple Load Balancing for Distributed Hash tables
-Torsha Banerjee
Presentation for Internet & Web Algorithms

2. Presentation Outline What is DHT (Distributed Hash Table)? Why DHTs?
An example
Applications
Load Balancing
How lookup works?
Paper Idea
Performance Evaluation
Conclusion
References

3. Distributed hash table Distributed application
What is DHT (Distributed Hash Table)?
Distributed hash table
Distributed application
get (key)
data
node ….
put(key, data)
No central server
Partitions an ID space among n servers
Each distributed server has partial list of where data is stored in the system
Keys are mapped uniformly to data values
A "lookup" algorithm is required to locate data
put (key, data) and get (key, data) functions are used to handle data
(Figure adopted from Frans Kaashoek)

4. What is DHT (Distributed Hash Table)? contd..
Architectures
CAN (Content Adressable Network) [1]
keys hashed into d dimensional space
operations performed are insert(key, value), retrieve(key)
chooses the neighbor nearest to the destination peer for storage
Chord [2]
maps key to a peer (server) for storage
maintains routing information as nodes join and leave the system

5. What is DHT (Distributed Hash Table)? contd..
Architectures contd..
PASTRY [3]
uses routing table, leaf set, neighborhood set, and file table
routes messages to the node with the closest ID to the key
low message overhead but higher latency and failure rate TAPESTRY [4]
the first DHT
routes messages to endpoints (peers)
objects are known by name, not location
creates a routing mesh of neighbors
each peer stores a neighbor map
complex and hard to maintain

6. Why DHTs?
Provides incremental scalability of throughput and data capacity as more nodes are added to the cluster
Robustness is achieved through data replication among multiple cluster nodes
Building block for peer-to-peer applications
Provides improved security and robustness

7. An Example Circular number space 0 to 127
Routing rule is to move clockwise until current node ID  key, and last hop peer (node) ID < key
Example: key = 42
Key is routed to node 58 in the clockwise direction
Node 58 “owns” keys in [34,58]
New entries always start with one known number
Newcomer searches for “self” in the network
hash key = newcomer’s node ID
Search results in a node in the vicinity where newcomer needs to be 13 58 81
111
127 33 24 97

8. Applications
Any application that requires a hash table (database systems, symbol table in compliers)
Storage, archival
Web serving, caching
Content distribution
Query & indexing
Naming systems
Communication primitives
Chat services
Application-layer multi-casting
Event notification services
Publish/subscribe systems

9. Load Balancing
Distributing processing and communications activity evenly across a computer network
Single point failure is avoided
Important for networks where it's difficult to predict the number of requests that will be issued to a server
Two components of load balancing
Consistent hashing over a 1-D space
Some kind of indexing topology is used to navigate the space

10. Load Balancing contd.. Consistent hashing:
Both keys and peers are hashed on a 1-D ring (e.g. Chord ring)
Keys are assigned to the nearest peer in the clockwise direction
Existence of overlay edges for faster search along the ring
Load imbalance may occur with increasing arc length
Maximum arc length is
Chord [2] solves this problem to some existent by introducing virtual peers

11. How Lookup works?
In an m-bit identifier space, there are 2m identifiers
Identifiers are ordered on an identifier circle modulo 2m
The identifier ring is called Chord ring
Key k is assigned to the first node whose identifier is equal to or follows k in the identifier space
Each node maintains a routing table with (at most) m entries
Finger table for node n is calculated from the adjacent table

12. How Lookup works? contd.. 10 [10,2) 7 [6,10) 6 5 [4,6) 4 [3,4) 3 1 1514 12 10 7 5 3 6 8 9 15 1
Example: Chord [2] 10
[10,2) 7
[6,10) 6 5
[4,6) 4
[3,4) 3
succ.
interval
start
Finger Table for Node 2
successor(6) = 7

13. Paper Idea Chord cannot totally solve the load balancing issue
One peer may be responsible for items
Number of edges per peer is in the worst case incurring higher messaging cost and storage space
Application of “power of two choices” [5,6] proposed in the paper [7]
Each peer represents a point in the circle (no finger table)
Peers arranged sequentially
d>=2 peers are chosen by an item for storage and the one with the least load is selected finally

14. Paper Idea contd..
0.4 1
Peer 1 is selected for storage of item x
The maximum load of any peer is
Four different hash functions are used to select for different peers
0.46 15
Item x- 14 2 3 12 5
0.6 10 6 7
0.51 9 8

15. Paper Idea contd..
If any two chosen peers have same load tie needs to be broken
Arbitrarily
Vocking’s scheme
Always-go-left scheme
Bins divided into d groups of size
Groups are numbered from 1 to d
If there are several locations with same load, the ball is placed in the location of the leftmost group
Choosing the least loaded arc of smallest length gives better results than Vocking’s scheme

16. Algorithm description
Paper Idea contd..
Algorithm description
h0 is used to map peers onto the ring
Hash functions h0(x), h1(x),…, hd(x) are first calculated by a peer to insert item x
Peers p1,…,pd are looked up in parallel corresponding to the hash functions h0(x), h1(x),…, hd(x)
pi with lowest load is selected for storing x
For searching an already inserted item x, again the d peers are looked up and the one storing the key-value pair is returned

17. Paper Idea contd..
Disadvantage is the increase in network traffic by an order of d
solved by the redirection pointers
All peers pj store a redirection pointer where pi is the least loaded peer
For searching x, hj(x) is used to find pj
If pj does not have x, it returns the redirection pointer
With uniform selection of hj, this extra step is required with probability
A particular peer needs to be active all the time
[8] solves this problem to some extent

18. Paper Idea contd.. Static placement of items leads to poor performance
Can be solve by letting item x change location when reinserted if pi, storing x previously, has become heavily loaded now
Schemes like load stealing and load-shedding are used
Load stealing- If pi is an under utilized peer it takes away load from other peers with heavier loads 2 14 12 10 7 5 3 6 8 9 15 1
0.9
0.02
Item x
Copy of Item x
Redirection pointer
Stealing of data

19. Paper Idea contd..
Load shedding- If pi is overloaded, it offloads data to lighter peers
Out of peers, p1,…,pd, x can be located in k least loaded peers thus replicating data
Enables parallel downloading from multiple sources
Item x 1 15
0.9 14 2
Redirection pointer 3 12
Offloading of data 5 10
0.02 6 7
Copy of Item x 9 8

20. Performance Evaluation
The three schemes compared with Chord [2] are
virtual peers
An unbounded number of virtual peers
The proposed power of two scheme
Simulation parameters
Number of peers
Number of items
1st percentile and 99th percentile load for Chord

21. Performance Evaluation contd..

22. Performance Evaluation contd..
Maximum load is key metric
The highest loaded peers have greater probability of failing
Cascading effect occurs leading to poor performance by neighboring peers
Performance of virtual peers is similar to that of Chord
Maximum load is high of the order of 350
Increase in topology maintenance traffic due to virtual peers
Load balancing is best in the unlimited case
Less variation in load compared to the “power of two” choice case
Maximum load is similar to “power of two” choice case of the order of 150

23. Performance Evaluation contd..
Unlimited resource scenario is unrealistic and expensive
The “power of two” choice case exhibits greater variation in load compared to the unlimited resource case
Maximum load is similar to the unlimited resource case
Better load balancing than virtual peer case
Routing information shared is less compared to Chord

24. Conclusion
The “power of two” choice case provides multiple storage options compared to Chord
Increased fault tolerance
Better performance in highly dynamic environments
Trade off is
small increase in amount of static storage at each peer
Small additive constant in search length

25. References
[1] Sylvia Ratnasamy, Paul Francis, Mark Handley, Richard Karp, Scott Shenker, “A Scalable Content-Addressable Network,” Proceedings of ACM SIGCOMM 2001.
[2] Ion Stoica, Robert Morris, David Karger, M. Frans Kaashoek, Hari Balakrishnan, “Chord: A Scalable Peer-to-Peer Lookup Service for Internet Applications,” Proceedings of the 2001 ACM SIGCOMM.
[3] A. Rowstron and P. Druschel, “Pastry: Scalable, distributed object location and routing for large-scale peer-to-peer systems,“ IFIP/ACM International Conference on Distributed Systems Platforms (Middleware), 2001.

26. References contd..
[4] Ben Y. Zhao, John Kubiatowicz, Anthony D. Joseph, “Tapestry: An Infrastructure for Fault-tolerant Wide-area Location and Routing,” SIGCOMM 2001.
[5] Azar, Y., Broder, A., Karlin, A., and Upfal, E., “Balanced allocations. SIAM Journal on Computing 29, 1 (1999),
[6] Mitzenmacher, M., Richa, A., and Sitaraman, R., “The Power of Two Choices: A Survey of Techniques and Results”, Kluwer Academic Publishers.
[7] John Byers, Jeffrey Considine and Mitzenmacher, “Simple Load Balancing for Distributed Hash Table”,