50 likes | 58 Views
This workshop presentation discusses the possible need for a new protocol or extensions to TCP in order to meet the requirements of high throughput and reliable bulk data transport for various users, including VLBI networks and particle physics experiments. It also explores alternative solutions, hardware assists, and the progress being made in protocol design.
E N D
TCP: Extensions or A New Protocol ?A User View Point Richard Hughes-Jones The University of Manchester Workshop on Protocols for Fast Long-Distance Networks Panel 2 PFLDNet Workshop February 2003 R. Hughes-Jones Manchester
Radio Astronomy: Very Long Base Line Interferometry • European VLBI Network includes Europe + China + Ukraine + South Africa • The amount of detail seen depends on the size of the telescope – hence VLBI • Sensitivity of the network depends on the rate at which the data recorded • Proof of Concept: 5 telescopes connected at 1 Gbit/s of user data this year. (10Gb) • Data Unusual: Does not matter is lose a small part so long as it is know it is lost • Other VLBI systems looking gat 30 Gbit/s per telescope (dark fiber / private SDH) PFLDNet Workshop February 2003 R. Hughes-Jones Manchester
Particle Physics • HEP experiments will need to move bulk data at high rate 24/24. • Planning on 2.5 Gbit/s going to 10 Gbit/s from Tier 0 to all large Tier 1 sites. • One LHC experiment will start trying to move 1 TByte/day of Monte Carlo data to CERN starting this November • OK its only 100 Mbit/s but there are 4 experiments and real data will be going in the other direction. • Other grid users of bulk data include: remote molecular visualisation, high definition TV for remote use of microscopes, tele medicine … PFLDNet Workshop February 2003 R. Hughes-Jones Manchester
Some “User” Requirements • High Throughput • Reliable bulk data transport: • bit-wise correct for many users • Minimum known loss for others • High throughput should start quickly – eg for http – TCP QuickStart ? • It should be fair: • To other users • Long RTT should not be penalised • Nor should MSS limit one* • Should be responsive to changes in Available bandwidth (both increase and decrease in rate) • So should be able to make efficient use of eg LBE QoS • Deployment – evolution or revolution or both? • General acceptance – IETF involvement • * Maybe think HW changes too PFLDNet Workshop February 2003 R. Hughes-Jones Manchester
Over to you … Some Points for Discussion • What’s the current work on new protocols? • What’s the investment likely to be ? – Remember OSI TP4 ?? • Can experience gained from NS and live tests of new stacks help choose the best way forward – I hope so ! • Alternatives to TCP end-to-end • UDP-based solutions eg Tsunami • Depot staging (cf BITNET) • iSCSI – this would be real disk-disk • … • Hardware Assists • What is the progress at the moment? • Can flexibility be included • RDMA • Can a protocol design make for very efficient HW assist? • Given a Lambda lightpath circuit and BERs, why run TCP on this? FEC OK? PFLDNet Workshop February 2003 R. Hughes-Jones Manchester