1. Challenges in the Next Generation Internet Xin Yuan Department of Computer Science Florida State University xyuan@cs.fsu.edu http://www.cs.fsu.edu/~xyuan

2. Internet Traffic doubles each year. Backbone traffic growth [Odlyzko01]: –100% per year from 1990 – 1994. –1000% per year in 1995 and 1996. –100% per year in 1997 through 2000. Internet is becoming more and more ubiquitous. More commercial applications. More multimedia applications.

3. The Next Generation Internet will meet all these challenges: Large bandwidth through the optical DWDM (Dense Wavelength Division Multiplexing) technology. multi-service. –service differentiation (Quality of Service). Security How to expose the large bandwidth to applications?

4. IP over WDM: protocol stack IP ATM SONET WDM IP WDM WDM-aware Electronic layer

5. IP over WDM: WDM-aware electronic layer: –Reconfiguration and load balancing. –Protection and restoration –Optical switching –Network management/control –Cross-layer optimizations

6. IP over WDM: Protection and restoration. –When a link is down, services should not be interrupted for too long. One optical link supports a large number of flows. SONET detects failure in 2-100us, restores operation in 60 ms. –Need to extend the SONET model to deal with the general mesh networks. –Determine the backup path together with the main path (more complex in routing). –How to utilize the capacity effectively?

7. IP Quality of Service: –To control the network service response so that is is predictable. –To allow the client to establish in advance the service response that will be obtained from the network. –To control the contention for network resources. –To allow for efficient total utilization of network resources.

8. IP Quality of Service: models –Integrated Service (IntServ): End to End QoS guarantee (What users want)? Too many states for each router to maintained (not scalable). Cannot be deployed incrementally. –Differentiated Services (DiffServ): Define PHB (per hop behavior). Each router distinguishes 64 queues (scalable). End-to-end QoS by provisioning? –Stateless QoS model: Carrying QoS information at each packet.

9. IP Quality of Service: other issues –Packet scheduling: FIFO does not work anymore. Fair queuing is quite expensive. –Routing: Taking QoS metrics into consideration can make it very difficult Scalable QoS routing for large networks can be hard.

10. Exposing hardware performance to applications. –Different applications result in different traffic patterns. Special communication optimizations can be performed by considering (1) the communication pattern and (2) the underlying network architecture. –Compiled communication for Message Passing Interface (MPI) programs over clusters of workstations.

11. Compiled Communication: –Traditional communication optimization: In the compiler: –Reducing the size and volume of communications. –Architecture independent optimizations. In the library: –Architecture dependent optimizations. –Don’t know the sequence of communications in applications. –Compiled communication: Compiler knows the sequence of communications in applications and the network architecture. Can perform architecture dependent optimizations across communication patterns.

12. MPI_Scatterv(…) … MPI_Scatterv(…) MPI_Open_Group(…) MPI_Data_Move(…) MPI_Close_Group(…) ….. MPI_Open_Group(…) MPI_Data_Move(…) MPI_Close_Group(…) MPI_Open_Group(…) MPI_Data_Move(…) ….. MPI_Data_Move(…) MPI_Close_Group(…)

13. Compiled Communication: How to make it happen? –Traditional communication libraries do not support this communication model (easy to use is the main goal). Compiled communication capable MPI library: –Should allow the compiler (user) to management network resources –In our prototype: »Compiler can manage multicast group. »Compiler can schedule the communications (for collective communications). –Need a compiler that can do the analysis.