1 Automatic Dynamic Run-time Optical Network Reservations John R. Lange Ananth I. Sundararaj and Peter A. Dinda Prescience Lab Department of Computer Science Northwestern University

2 Introduction Recently the Grid community has begun turning the network into a standard grid service –Network and compute resources are still very different –Requires in depth understanding of program behavior or targeted application development VRESERVE –Enable unmodified applications to efficiently utilize reservation based networks without user intervention

3 Overview Reservations and Circuit Switching Optical Networking –OMNInet and ODIN Virtual Machines –Virtuoso Overlay Networking –VNET and VTTIF Putting it all together –VRESERVE Performance Evaluation

4 Reservable Networks Generic Reservation API –CreatePath(,,, ); –TeardownPath(, ); Somebody or something has to make the API calls Circuit switching and network reservations are inherently linked –Establishing a circuit is similar to reserving the network elements along a path

5 Reservation Issues Parameter Values –How does someone determine the required bandwidth and latency? –Are the endpoints always evident? Highly parallel applications in a grid environment Migratable Virtual Machines Time –What is the flow duration? –What is the traffic behavior pattern? Applications often have cyclical behaviors (compute and send)

6 Reservation Problems Complexity: –Someone has to call the reservation API Developer must incorporate API calls into application User must reserve network on application’s behalf Efficiency: –Efficient reservations require understanding application’s behavior and topology –Networks usually reserved in spatial and temporal blocks Networks reserved to connect all hosts Networks reserved for entire execution time

7 Optical Networks Recent advances in switching technology have brought renewed interest to the area Offer provisioning of dynamic lightpaths –One of the latest incarnations of circuit switched networking –Emergence of network reservation grid services –Generating a large deal of interest in the high performance computing and grid communities NLR, Canarie, Netherlight OptIPuter

8 OMNInet and ODIN OMNInet –Experimental dynamically configurable optical network –Provisionable wavelengths (lambdas) –All-to-All topology –Connects research centers in Chicago and Northern Illinois ODIN –Reservation system for OMNInet Developed by iCAIR –International Center for Advanced Internet Research –Mambretti, J., Weinberger, J., Chen, J., Bacon, E., Yeh, F., Lillethun, D., Grossman, B., Gu, Y., and Mazzuco, M. The photonic terastream: Enabling next generation applications through intelligent optical networking at iGRID2002.

9 ODIN Lightpath Reservation System –Implements path discovery based on user supplied endpoints –Also beginning to look at scheduling services –Command line client interfacing with a trusted server Interface –oclient -c Path Creation –oclient -t Path Teardown

10 10 GE Lake Shore Photonic Node S. Federal Photonic Node W Taylor 1 Gbps 10 GE Optera Gb/s TSPR Photonic Node  PP GE PP 8600        Optera Gb/s TSPR Optera Gb/s TSPR     10 GE PP 8600 NWUEN-2 NWUEN-3 NWUEN-4 NWUEN-8NWUEN OFA Optera 5200 OFA 5200 OFA VM Network Path Taken Potential Paths Fiber Host Internet

11 Virtuoso Virtual Machine marketplace Collection of remotely distributed VMs that appear to reside on the same LAN Provides many opportunities for optimization –Adaptive overlay networks, VM migration, resource scheduling, etc… –A. Sundararaj, A. Gupta, and P. Dinda, Increasing Application Performance In Virtual Environments Through Run-time Inference and Adaptation HPDC 2005 –A. Sundararaj, M. Sanghi, J. Lange, P. Dinda, An Optimization Problem in Adaptive Virtual Environments (MAMA 2005), To Appear

12 Application Data Control

13 VNET Virtual Networking for Virtual Machines Allows VMs to behave as if they were on the same LAN –All remote VMs tunnel traffic to a central proxy Default star topology A. Sundararaj, P. Dinda, Towards Virtual Networks for Virtual Machine Grid Computing –USENIX VM 2004

14 Application Data Control

15 VTTIF Application Topology Inference Extracts network topologies from application behavior Generates a global traffic matrix defining the global application network topology A. Gupta, P. Dinda, Inferring the Topology and Traffic Load of Parallel Programs Running In a Virtual Machine Environment –Workshop on Job Scheduling Policies for Parallel Processing, 2004

16 VADAPT Overlay Network Adaptation Modify virtual network to match actual network topology –Creates overlay links between communicating VMs Based on global reduction of VTTIF matrices A. Sundararaj, A. Gupta, P. Dinda, Dynamic Topology Adaptation In Virtual Networks of Virtual Machines –LCR 2004 A. Sundararaj, A. Gupta, and P. Dinda, Increasing Application Performance In Virtual Environments Through Run-time Inference and Adaptation –HPDC 2005

17 Application Data Control

18 VRESERVE Network Reservations for Overlay Networks Extension to VADAPT –Allows true adaptation of the network Components –Reservation API interface ODIN CLI –Routing mechanism Address mapping service –Path Discovery Depends on Reservation System functionality Allows for scheduled and delayed reservations –Greatly increases usability for scheduler based reservation systems

19 Optical Overlays VRESERVE routing is accomplished with VNET overlay links Issues –Reservation based resources usually aimed at high performance –Application unaware of changing network conditions TCP performance is typically poor in high performance networks Benefits –Application unaware of changing network conditions –Routing is much easier at the overlay level Network is capable of reacting to global state changes

20 10 GE Lake Shore Photonic Node S. Federal Photonic Node W Taylor 1 Gbps 10 GE Optera Gb/s TSPR Photonic Node  PP GE PP 8600        Optera Gb/s TSPR Optera Gb/s TSPR     10 GE PP 8600 NWUEN-2 NWUEN-3 NWUEN-4 NWUEN-8NWUEN OFA Optera 5200 OFA 5200 OFA VM Fiber (MWUEN-4) length = 5 miles Host Internet VNET VTTIF + VADAPT VRESERVE ODIN

21 Evaluation This is an existence proof –Possible to automatically reserve network resources on unmodified application’s behalf Caveats and Disclaimers –Scalability is unknown –Final performance measurements are also unknown Infrastructure issues prevented complete performance evaluations –Only one functional optical link –Network dismantled during experiments

22 Optical Network Performance Latency (ms) Optical.321 Internet.498

23 Initial VNET performance (optical Network)

24 Performance Improvement (Synthetic BSP benchmark) Demonstrable improvement by using automatic reservations

25 Making VNET faster Overlay links switched to UDP –Reliable transport implemented in VM TCP stack –2x improvement Lookup table caching –3x improvement Future: –Memory mappings for packet operations –In kernel forwarding –Specialized VM device drivers

26 VNET Performance (Round 2) No results on Optical Network –Infrastructure unavailable Experiments conducted between cluster nodes w/ gigabit Ethernet VNET performance only –Reservation and adaptation systems unused Overheads insignificant

27 VNET Performance (Gigabit switch)

28 Future Work Where does this fit into the overall optimization problem? How do we define the overall optimization problem? Extension to other reservation systems, e.g. GARA

29 Conclusion Networks are not generic grid services Middleware is required to make reservation networks usable VRESERVE –Enable unmodified applications to effectively utilize reservation based networks without user intervention

30 Prescience Lab – Virtuoso – DOT (Distributed Optical Testbed) –

31 Optical Network Performance Throughput (MB/sec) Optical TTCP Optical SCP 11.5 Internet TTCP 11.2 Internet SCP 10.4 Latency (ms) Optical Network.321 Internet.498 Throughput