Data Transport Challenges for e-VLBI Julianne S.O. Sansa* * With Arpad Szomoru, Thijs van der Hulst & Mike Garret
27 October Outline Network performance tests Simulation results conclusion
27 October Network Performance Measurements Investigate critically several connections established. Wire speeds suggests much higher throughput than what application data realises. TCP Congestion Control algorithm (AIMD) –SSACK:Cwnd Cwnd +1 –CAACK:Cwnd Cwnd + 1/Cwnd DROP: Cwnd Cwnd -1/2*Cwnd Cwnd = max. # packets that TCP injects into network before receiving ACK. Cwnd optimal ~ Throughput *RTT Cwnd average = 1.22*MSS/sqrt (p) [Floyd & Fall (1999), Padhya et.al (1998)]
27 October Specific Questions How much bandwidth is available to the these TCP connections? Is it what is seen by the app? If it is less than the theoretic available b/w, what is the bottleneck? How do we minimise this bottleneck? How do multiple TCP connections share available bandwidth? What is the stability of these TCP connection (repeatability /predectability)?
27 October Results with web100 File transfer of 10 GB & 1GB file Modified file transfer (app socket buffers) Memory-memory with iperf
27 October Cwnd, RwinRcvd & for a file transfer / memory-memory
27 October Achieved/Available throughput
27 October Summary Test results Memory –MemoryFile transferModified file transfer Disk2net-net2file (yet to be done) Cwnd (bytes) TCP (Mbps) UDP (Mbps) Cwnd (bytes) TCP (Mbps) Cwnd (bytes) TCP (Mbps) CwndTCPUDP Bench via Amste rdam
27 October NIC RTT/loss discrepancies
27 October The bottlenecks Application socket buffers Hardware (PCI bus limit, NICs) The OS (more or less tuned optimally) The transport protocol (TCP) –Window limits –Retransmissions –Interface stalls –Vendor specific implementations (Other Reductions)
27 October Transport Protocol Analysis Already many proposals to alter this behaviour: HighSpeed TCP, scalable TCP, Westwood TCP, HTCP, Vegas, FAST, BIC, C-TCP
27 October Loss-based, delay-based,or equation-based? Which way do we go? Consider getting the best out each world/Allow the application to dynamically detect network conditions & decide which algorithm to use.
27 October Preliminary Simulation results Simulated file transfer of bench via Amsterdam scenario TCP UDPHSTCPFAST Cwnd (bytes) 1, n/a4, T/put (bps) 61,600,000960,000,00061,600,000310,870,560
27 October Cwnd for the simulated protocols
27 October Achieved Throughput for the simulated protocols
27 October Conclusions & further work Hardware (PCI bus, NICs,) on end systems as well as the application (buffers) need to be optimised. Model TCP data flows & relate flow analysis with correlation. More simulation work on Transport Protocol analysis (response function)
27 October References Floyd & Fall (1999) “Promoting the use of end-to- end congestion control in the internet”, IEEE/ ACM Trans. on Networking, August Padhya et.al (1998) “Modeling TCP throughput: A Simple model and its empirical validation” in Proc ACM SigCOMM 1998 Antony et.al(2004) “Exploring Practical Limitations of TCP over Transatlantic Networks” submitted Elsevier Science(2004)