Presentation is loading. Please wait.

Presentation is loading. Please wait.

Announcements Your homework is due on September 19 th. Your homework is due on September 19 th. I will be away starting Sept 5 th.

Similar presentations


Presentation on theme: "Announcements Your homework is due on September 19 th. Your homework is due on September 19 th. I will be away starting Sept 5 th."— Presentation transcript:

1 http://dblab.usc.edu Announcements Your homework is due on September 19 th. Your homework is due on September 19 th. I will be away starting Sept 5 th and will see you on Sept 19 th. I will be away starting Sept 5 th and will see you on Sept 19 th.  Make sure to attend lectures and listen to video taped tutorials. They are essential for your homework and project. Professor Francois Bar is guest lecturer on September 26 th. Professor Francois Bar is guest lecturer on September 26 th. Professor Anne Friedberg is guest lecturer on October 10 th. Professor Anne Friedberg is guest lecturer on October 10 th.

2 http://dblab.usc.edu Any Homework Questions?

3 http://dblab.usc.edu Question from last week

4 CSCI 599: Replication in Ad Hoc Network of Wireless Devices Shahram Ghandeharizadeh Computer Science Department University of Southern California

5 http://dblab.usc.edu Outline Target environment: H2O networks. Target environment: H2O networks. Definition of access cost. Definition of access cost. GreedyDual replacement technique. GreedyDual replacement technique. Conclusion and future research. Conclusion and future research.

6 http://dblab.usc.edu Target Environment Key characteristics: Key characteristics:  Limited by the bandwidth and radio range of wireless devices (less constrained by energy and mobility).  Management of devices is decentralized.  Each device is autonomous.  Devices cooperate when in radio-range of one another.

7 http://dblab.usc.edu Home-to-Home Online (H2O) devices collaborate to deliver continuous media: Home-to-Home Online (H2O) devices collaborate to deliver continuous media: A H2O device is a wireless device with a powerful processor and abundant amount of storage. A H2O device is a wireless device with a powerful processor and abundant amount of storage. Example: H2O Clouds S. Ghandeharizadeh, H2O Clouds: Issues, Challenges and Solutions, in IEEE Pacific-Rim Conference on Multimedia, 2003.

8 http://dblab.usc.edu Cost of a clip Euclidean distance between the nearest replica and the requesting node. Euclidean distance between the nearest replica and the requesting node. In a grid, the Euclidean distance between two nodes is defined as: In a grid, the Euclidean distance between two nodes is defined as:

9 http://dblab.usc.edu Euclidean distance Euclidean distance between the orange and green nodes is: Euclidean distance between the orange and green nodes is:  2 2 + 3 2 = 4 + 9 = 13  SQRT (13) = 3.61

10 http://dblab.usc.edu Euclidean distance: 1 replica Maximum Euclidean distance to the green node containing a replica of the required data item is: Maximum Euclidean distance to the green node containing a replica of the required data item is:  2 2 + 2 2 = 4 + 4 = 8  SQRT (8) = 2.83

11 http://dblab.usc.edu Euclidean distance: 4 replicas Maximum Euclidean distance to the green node containing a replica of the required data item is: Maximum Euclidean distance to the green node containing a replica of the required data item is:  1 2 + 1 2 = 1 + 1 = 2  SQRT (2) = 1.41 Four fold increase in number of replicas provided only a factor of 2 improvement in Euclidean distance. Four fold increase in number of replicas provided only a factor of 2 improvement in Euclidean distance.

12 http://dblab.usc.edu Main assumption 1. Claim: In a 2-D mesh network, the Euclidean distance between the nearest replica and the requesting node is:  d i is density of replicas for the i-th object.  Can you provide a proof for this? (Potential project for the theoreticians in the class.) 2. Ignores issues risen from the boundary by assuming an unbounded network.

13 http://dblab.usc.edu GreedyDual Algorithm L = 0 For each request for object i do If i is in the local cache then H(i) = L + C(i) Else while there is no space for i do L = min{ H(j) such that j is in cache } Evict j which satisfies H(j) = L Retrieve and store i H(i) = L + C(j) When object size is variable then replace C(i) with C(i)/object-size. When object size is variable then replace C(i) with C(i)/object-size.

14 http://dblab.usc.edu GreedyDual Algorithm Consider a node with storage capacity of 3 objects. Its storage is occupied with objects A, B, and D Consider a node with storage capacity of 3 objects. Its storage is occupied with objects A, B, and D Cost of objects is as follows: Cost of objects is as follows:  C(A) = 12, C(B)=7, C(E)= 5, C(D)=1 Reference for object E causes the algorithm to evict D because it has the lowest cost, L is set to 1, H(E)=6. Reference for object E causes the algorithm to evict D because it has the lowest cost, L is set to 1, H(E)=6. If the next reference is again for E then H(E)=? If the next reference is again for E then H(E)=?

15 http://dblab.usc.edu GreedyDual Algorithm Consider a node with storage capacity of 3 objects. Its storage is occupied with objects A, B, and D Consider a node with storage capacity of 3 objects. Its storage is occupied with objects A, B, and D Cost of objects is as follows: Cost of objects is as follows:  C(A) = 12, C(B)=7, C(E)= 5, C(D)=1 Reference for object E causes the algorithm to evict D because it has the lowest cost, L is set to 1, H(E)=6. Reference for object E causes the algorithm to evict D because it has the lowest cost, L is set to 1, H(E)=6. If the next reference is again for E then H(E)=7 If the next reference is again for E then H(E)=7 What object is evicted if D is referenced? What object is evicted if D is referenced?

16 http://dblab.usc.edu GreedyDual Algorithm Consider a node with storage capacity of 3 objects. Its storage is occupied with objects A, B, and D Consider a node with storage capacity of 3 objects. Its storage is occupied with objects A, B, and D Cost of objects is as follows: Cost of objects is as follows:  C(A) = 12, C(B)=7, C(E)= 5, C(D)=1 Reference for object E causes the algorithm to evict D because it has the lowest cost, L is set to 1, H(E)=6. Reference for object E causes the algorithm to evict D because it has the lowest cost, L is set to 1, H(E)=6. If the next reference is again for E then H(E)=12 If the next reference is again for E then H(E)=12 What object is evicted if D is referenced? B, L is now 7, H(D) = 8 What object is evicted if D is referenced? B, L is now 7, H(D) = 8

17 http://dblab.usc.edu Modified GreedyDual Change the algorithm to not increase H(i) when it is referenced frequently: Change the algorithm to not increase H(i) when it is referenced frequently: L = 0 For each request for object i do If i is in the local cache then do nothing Else while there is no space for i do L = min{ H(j) such that j is in cache } Evict j which satisfies H(j) = L Retrieve and store i H(i) = L + C(j)

18 http://dblab.usc.edu Evaluation Assumes a fixed square area. Assumes a fixed square area. 40,000 nodes in the mesh network, distributed randomly. 40,000 nodes in the mesh network, distributed randomly. A node may store 25 objects. A node may store 25 objects. Number of objects is 10,000 with Zipf distribution of access. Number of objects is 10,000 with Zipf distribution of access. Dijkstra’s shortest path algorithm Dijkstra’s shortest path algorithm  Link cost = physical distance On average a node has more than 5 directly connected neighbors. On average a node has more than 5 directly connected neighbors. Normalized access cost = cost / diagonal of the square area. Normalized access cost = cost / diagonal of the square area.

19 http://dblab.usc.edu Performance Results More skewed

20 http://dblab.usc.edu Performance Results

21 http://dblab.usc.edu Performance results What other number might be reported? What other number might be reported?

22 http://dblab.usc.edu Performance results What other number might be reported? What other number might be reported?  Number of replacements performed by each node.  Percentage hit ratio for each node: average & standard deviation.  Cold start, and finish Was a node thrashing? Was a node thrashing?  Is it possible for a node to thrash?

23 http://dblab.usc.edu Possible project 1 A literature survey of dynamic replacement policies for H2O networks. It must include a comparison with: A literature survey of dynamic replacement policies for H2O networks. It must include a comparison with:  Memory management techniques for an audio/video-on-demand server.  Cache servers designed for the Internet.  Cache servers designed for environments such as Akamai.

24 http://dblab.usc.edu Possible project 2 Compare GreedyDual with a technique such as LRU: Compare GreedyDual with a technique such as LRU:  How much improvement does GreedyDual provide?  How does this improvement change as a function of storage capacity of a node?  Does one technique exhibit a thrashing behavior? How would you set the experiment up? How would you set the experiment up? What would be the important parameters to quantify and analyze? What would be the important parameters to quantify and analyze? Remember the 10 Commandments when attempting this project. Remember the 10 Commandments when attempting this project.

25 http://dblab.usc.edu Class Collaborative Project Theme: What if cars became driver-less by the year 2100? Theme: What if cars became driver-less by the year 2100?  What would information services would be available in the car?  How would the car be used as an extension of a home/office? Deliverable: Deliverable:  A script that serves as the basis of a video-clip (similar to AutoMatas) describing the future information service in cars that drive themselves.  What signs are there that cars will be able to drive themselves?


Download ppt "Announcements Your homework is due on September 19 th. Your homework is due on September 19 th. I will be away starting Sept 5 th."

Similar presentations


Ads by Google