/  4
 
 
PERFORMANCE EVALUATION OF PEER-TO-PEER APPROACH TORESOURCE DISCOVERY FOR LARGE SCALE GRID ENVIRONMENTS
Sivadon Chaisiri
1
and Putchong Uthayopas
2
 
ABSTRACT:
Grid environment usually consists of large number of shared resources distributedacross multiple locations. Hence, the problem of resource discovery can be profoundly complex dueto the size and complexity of grid system. Recently, peer-to-peer system has been widely used due toits massive resource sharing environment. The merging of the capability of both systems has apotential to create a very large scale sharing environment with rich functionality. In this work, thearchitecture for building grid networks based on peer-to-peer approach is proposed. Furthermore,some resource discovery algorithms based on the architecture are explored, and evaluated by buildinga simulation. The results of this work will lead to the development of the effectively large scale gridnetwork in the future.
 KEYWORDS:
peer-to-peer computing, grid computing, resource discovery algorithm
1.
 
INTRODUCTION
Grid environment consists of a large set of resources, which are distributed among geographicallocations (Foster 1999) (Foster 2001). In large set of shared resources, a basic problem is locatingresources when they are represented only by a set of desired attributes (i.e. a Unix computer withphysical memory of 256 MB) rather than a globally unique identifier (such Internet address).Moreover, it is impossible to assign the unique name to resources that change their attributesdynamically (e.g. CPU load and available memory). The complexity and size of grid environmentseven make it more challenging to discover the preferred resources in an appropriate amount of time.Recently, peer-to-peer system is emerging as a new sharing mechanism that widely used around theworld (Alfred 2003). Millions of resources are shared among massive number of computers that canfreely join or leave the system. The dynamism and scale of peer-to-peer system seems to be what isneeded by the grid system. Therefore, the convergence of both systems (peer-to-peer and grid) canpotentially create a large distributed system that is scalable and efficient. Nevertheless, the problem of discovering resources under this system is still existed.In this work, a proposed architecture for that combines the benefit of grid and peer-to-peer system arepresented. Furthermore, several resource discovery algorithms also are proposed for this architecture.Then, the proposed architecture is simulated to evaluate the effectiveness of these proposedalgorithms.
2.
 
PROPOSED ARCHITECTURE
In this work, resources (e.g. computers, printers, sensors etc.) that connected together using thenetwork are termed “
 peer 
”. There are 4 types of peer as follows.
1
Master Student ,High Performance Computing and Networking Center, Kasetsart University
2
Assistant Professor, Department of Computer Engineering, Kasetsart University
 
 
 
Edge Peer
is a peer that shares resources to other peers.
 
 
Super Peer
is a peer that keeps a set of locations of edge peers and known super peers.Furthermore, it has a duty to propagate queries (for resources) received from any peers to itsknown peers. We call peers that known by other as “
neighbored peers
”.
 
 
VO Peer
is a peer that forms a virtual organization. Moreover, it defines the policy formanaging peers inside the VO. To join a specific VO by any peers, they need to join viaregistering with the VO peer that manages the VO.
 
Global Peer
is a peer that provides the naming directory of VO peers that need to advertisetheir existence to other peers that need to join with.Grid system proposes the term “virtual organization (VO)” for managing shared resources based onthe defined policy. To merge the capability of peer-to-peer to the grid system for building asystematic sharing environment, we must provide the concept of virtual organization as well.However, the conventional peer-to-peer systems do not provide the facility for defining global policy.
 
Although, some literature such as (Iamnitchi 2001) has also proposed the approach and structure formerging both grid and peer-to-peer systems, the building of VO has not fully addressed. Our work isdifferent from since we propose the components for building the grid environment that also addressthe problem of building a virtual organization as well. In addition, we show that the proposealgorithms can easily scale to as many as 100,000 peers while other works usually address the systemof only up to 5,000 peers. In the next section, several algorithms for resources discovery will beproposed.
3.
 
ALGORITHMS
Most of the resource discovery algorithms (included our proposed algorithms) for peer-to-peersystems are based on the widely used flooding strategy. This strategy is to flood queries from one peerto its neighbored peers. Firstly, a user submits queries from his peer (called a requester) to otherneighbored peers. When a neighbored peer receives the queries, it will process the queries over itsown local shared resource. If any results are found, it will send a response message (includedinformation for accessing to the resources) back to requester. The neighbored peer also forwards thequeries to its own neighbored peers. To prevent infinite loop of flowing queries in the network, everymessage of query is handled by
time to live
(TTL) that specifies how many hops the message maytake. When a peer receives a query, it decrements the TTL, and if the TTL is greater than 0, it stillforwards the query to all its neighbored peers. Until the TTL is 0, the query is stopped forwarding.The advantage of this flooding strategy lies on the fact that it does not require any centralized servers.However, this strategy produces large amount of messages in order to propagate the queries to peers.This it will degrade the overall performance of network severely.In this work, 3 resource discovery algorithms are proposed as follows:
 
 Algorithm-1
 
(named “Query-Filter Algorithm)
this algorithm is designed to improve the oldflooding algorithm that we labeled as
 Algorithm-0 (named “Dumb Flooding Algorithm”)
.
 
The query-filter algorithm improves the dumb flooding algorithm by reducing the bursting of produced messages. Each query attaches the information about routed peers (the peers thatrouted by queries). That is, when a query passes each peer, the query will attach informationto inform that it has already passed this peer. Consequently, before propagating queries toothers by a peer, it must check the routed peers from the information. If it found that the
 
query had to pass their neighbored peers, they will not propagate it to them; otherwise, theycontinuously flood it to them. Furthermore, every peer also filters seen queries by checkingfrom the identifier number of every query. That is, each peer must store identifier numbers of received queries. If it found the seen queries, it will not propagate them to others.
 
 Algorithm-2
 
(named “Backing Links Algorithm”)
this algorithm improves on the query-filter algorithm. The algorithm is done when a preferred resource is found, the responsemessage that included links (URLs for connecting peers) of all routed peers will be returnedto each others (included the requester who send the queries) for connecting them asneighbored peers. We assumed that the requester should discover preferred resources faster if it knows a lot of neighbor peers.
 
 Algorithm-3
 
(named “Backing Resource Links Algorithm”)
this algorithm improves thebacking links algorithm by inserting additional information to the response message. Theadditional information is about the method for accessing the preferred resource. We assumedthat if peers know about resources of many peers, the performance of resource discoveryshould be speeded up.
4.
 
SIMULATION AND EVALUATION
We build the simulated environment for evaluate the proposed resource discovery algorithms. Thesimulation is implemented by HyperSim (Sugree 2003). We simulated 4 sharing environmentscomposed with different numbers of peers: 100, 1,000, 10,000, and 100,000 peers. In eachenvironment, every VO consists of one selected VO peer to manage a policy. Each peer shares onlyone unique resource. Every peer is assumed to be connected via 1 Mbps network.
Table 1. Configuration of Simulated EnvironmentsEnvironmentNo.Number of PeersNumber of VONumber of Super peers/VONumber of Peers/VO
1 100 1 1 1002 1,000 2 2 5003 10,000 4 4 2,5004 100,000 8 8 12,500
We have 2 metrics for testing the algorithms. The first metric is the number of produced messagesoriginated from flooding queries to peers. For testing this metric, we simulate situations that there is apeer submits a query for finding out a preferred resource. We observe the number of producedmessages within 10 seconds by setting TTL = 12 hops. From the figure 1., we test two algorithms(dumb flooding and query-filter algorithms) and redo this testing for 100 times on each environment.Then we calculate the average of the number of produced messages. We found that the number of peers will dramatically affect to the message production, especially produced by the dumb floodingalgorithm. But our three proposed algorithms reduce tons of produced messages.
Figure 1: Number of messages produced by thealgorithms simulated on each environmentFigure 2: Average response time taken by thealgorithms simulated on each environment
 
050000010000001500000200000025000003000000100 1000 10000 100000
Number of peers
   N   u   m   b   e   r   o   f   p   r   o   d   u   c   e   d   m   e   s   s   a   g   e   s
Algorithm-0Algorithm-1Algorithm-2Algorithm-3
010002000300040005000600070008000900010000100 1000 10000 100000
Number of peers
   A  v  e  r  a  g  e   R  e  s  p  o  n  s  e   T   i  m  e   (  m  s .   )
Algorithm-0Algorithm-1Algorithm-2Algorithm-3

Share & Embed

More from this user

Add a Comment

Characters: ...