NETWORKED EMBEDDED SYSTEMS

1. NETWORKED EMBEDDED SYSTEMS SRIKANTH SUBRAMANIAN

2. Agenda • Overview Networked Embedded Systems (NES) • NES built on ASIPs • NES built on General Purpose Processors • Quantitative performance comparison of various NES for TCP and UDP • Online HW/SW partitioning of NES

3. Overview • NES are employed in Devices that form the backbone of communication networks • Routers, Network Bridges (Switches), Telephone Switches etc • Perform the task of data processing, Network Connectivity and Service delivery

4. Overview • NES were mostly implemented on Single Purpose Processors (SPPs) • SPPs - No Flexibility • Pose problems when the requirements change • Alternatives : General purpose processors or ASIPs (Network Processors)

5. Network Processors • ICs specifically built for networking application • Software Programmable Devices • Optimized features for networking applications : pattern Matching, Queue management, Data bit field manipulation etc.

6. NES based on ASIPs • Stareast: • Consists of a Baseboard and two daughter boards • Baseboard contains Intel IXP425 (533MHz) Processor

7. Scalable performance, Reduced power consumption, Low cost Deliver a range of data, voice, security and I/O features Distributed processing Architecture Combination of Intel XScale (an ARM Processor) and 3 Network Processing Engines (NPEs) Intel IXP425 Processor

8. Intel IXP425 Processor • XScale - Control plane • NPEs - computationally intensive data • Parallel Operation of Xscale and NPEs

9. NES based on ASIPs • Netgear WAG302: • Most commonly used wireless Access Point • Based on Intel IXP422B processor (266 MHz)

10. NES based on GPPs • Soekris net4826-50 • Based on AMD Geode Processor (266 MHz)

11. NES based on GPPs • Used to create fully customized routers and access points • Low cost, advanced communication features • AMD Geode processor comes under the X86 processor family

12. Quantitative Performance Comparison • Objectives: • Performance comparison Between NES based on GPP and NES based on ASIP • Performance comparison between Two NESs based on ASIP with one running on a commercial Operating system and another running Open source operating system

13. Experimental Setup • Three Stareast boards: • One running Montavista 4.0 OS • The other two running Snapgear versions 3.1 and 3.3 • Netgear WAG302 Running openWRT • Soekris net 4826 running Voyage Linux (Debian) distribution for X86 processors

14. Experimental Setup

15. Experimental Setup • To Study the behavior of the NESs, D-ITG traffic generator is used • Can generate IPv4/IPv6 traffic replicating the appropriate stochastic processes for both IDT( Inter Departure time) and PS (packet size) • Collect Statistics of Quality of Service (QOS) parameters: Throughput, Jitter, Packet loss and Delay (Latency)

16. Experimental Analysis • NES boards are connected back to back with the Workstation • Testing is performed using both TCP and UDP in the transport layer • Two types of tests are performed: • Discover the number of packets per second the devices are able to generate for fixed packet size • Measure bit rates, jitter, packet loss for different packet rates and sizes

17. Results • Packet Rate:

18. Results • Bitrate for UDP:

19. Results • Bitrate for TCP:

20. Results • Jitter for UDP:

21. Results • Jitter for TCP:

22. Results • Packet Loss for UDP:

23. Conclusions • NES based on ASIP running a commercial OS provides better performance as compared a NES based on GPP running a commercial OS • NES based on ASIP running open source OS are still less efficient as compared to commercial OS • Hence NES using network processors can play a major role in data intensive real time applications.

24. Online HW/SW Partitioning • Need: • Optimal partitioning of load into HW and SW during compile time is difficult • Arrival of new tasks during execution time • Failure of a node during run time

25. Network Architecture • Graph Theory: • Structure of Network ga = (N,C)

26. 3 -1 -1 -1 -1 3 -1 -1 -1 -1 3 -1 -1 -1 -1 3 Network Architecture • Laplacian Matrix: Given a graph G with n vertices (without loops or multiple edges), its Laplacian matrix is defined as

27. Network Architecture • Assumptions w.r.t Network • Architecture graph is undirected • each computational task may be assigned to each node in the network without restriction • HW Reconfiguration • Temporal partition: Temporal HW/SW partition at time t is an assignment of each task pЄ P(t) to a resource N ( t ) as well as the indication whether the task is implemented in HW or SW.

28. Network Architecture • Workload characterization: Each task pjЄ P(t) causes a unique load whj on resource niЄ N ( t ) if implemented in HW and a load of wsj, if implemented in SW. • Load: For HW - The fraction of total area occupied by the load For SW - The fraction of execution time and period.

29. Diffusion Algorithm • load exchanges between two adjacent nodes are determined in each iteration as: yk-1c = β(wk-1i – wk-1j) for all c = {ni, nj} Є C wki = wk-1i - ∑ yk-1c c = {ni, nj} Є C • Changing β in each iteration k has shown that the convergence speed can be drastically improved to exactly m - 1 iterations Choosing β = 1/גk where 1 ≤ k ≤ m – 1 m – Number of distinct Eigen Values of the Laplacian matrix

30. Optimization Flow

31. Optimization Flow • Objectives: • Find a bi-partition such that the load is balanced between HW and SW i.e.; minimize ∑|N| wSi - ∑|N| wHi i = 1i = 1 • Effective Load Balance i.e.; minimize |w’ – max{maxi:niЄN{wSi }, max{maxi:niЄN{wHi }| by using Evolutionary Algorithm applied to encode implementation selection • Diffusion Algorithm only balances load between nodes not HW/SW load

32. Discrete Diffusion Algorithm • Need: It is advisable not to split one process and distribute it to multiple nodes. This increases the data traffic in the network. Let ycontkcbe the real-valued continuous and ydisckcthe discrete flow on one edge c in iteration k such thatydisckc doesn’t exceedycontkc ydisckc ≤ ycontkc +ek-1c with e0c = 0 ekc =ycontkc +ek-1c - ydisckc for all c={ni, nj} Є C An additional adjustment step is introduced emc =em-1c - yadjc

33. Experimental Analysis • Evaluation the discrete diffusion algorithm for different types of a network like meshes with 3x3 or 4x4 nodes, a ring and a chordal ring with 8 nodes • In the beginning all tasks are mapped onto a single resource node • Focus is set on the load error |w’ – wi| and the congestion in the network

34. Experimental Analysis

35. Experimental Analysis

36. Thank You Questions?