LCN 2006: 31st IEEE Conference on Local Computer Networks, Tampa, FL, USA, November 2006 Considerations of SCTP Retransmission Delays for Thin Streams Jon Pedersen1, Carsten Griwodz1,2 & Pål Halvorsen1,2 1Department of Informatics, University of Oslo, Norway 2Simula Research Laboratory, Norway {jonped, griff, paalh}@ifi.uio.no Overview Latency problems for thin streams SCTP as an alternative to TCP Experiments New experiments Conclusions LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Thins Streams Transport protocols being developed for throughput-bound applications BUT, there exist several low-rate, time-dependent applications Anarchy Online MMORPG Case Study average delay: max delay: packets per second: average packet size: average bandwidth requirement: LCN 2006, Tampa, FL, USA, November 2006 ~250 ms 67 seconds (6 retransmissions) <4 (less then one per RTT) ~120 bytes ~4 Kbps 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Stream Control Transmission Protocol sender SCTP should support signaling “require response between 500 – 1200 ms” … or “initiation of error procedures” [rfc 2719] Network SACK (re)transmission queue acknowledged error-free transfers data fragmentation according to MTU packet boundary maintenance sequenced delivery within multiple streams bundling partial reliability … suppose to address low latencies receiver LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Test Set Up Linux 2.6.15 with lksctp 100 bytes packets 4 packets per second 3.2 Kbps SCTP LCN 2006, Tampa, FL, USA, November 2006 Network emulated using netem • dropp • delays (RTTs: 0, 100, 200, 400 ms) 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Results: lksctp for Thins Streams Even worse than TCP!!! Why these high delays? Two ways of triggering retransmissions of a lost chunk… LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Retransmission by Time-Out sender Timeout is dependent on minRTO = 1000 ms estimated RTT based on SACKs BUT SACKs are delayed o o one ACK for two packets or 200 ms timer influences estimated RTT, for thin streams retransmissionespecially of packet with green chunks due to timeout RTO value grows Network SACKSACK (re)transmission queue receiver LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Retransmission by Fast Retransmit sender receiver 4 SACKs needed for fast retransmit + thin streams = “all” retransmissions due to timeouts no no no no SACK SACK SACK SACK Network LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Enhancement: Removal of Exponential Backoff time in RTTS sender receiver 8 6 4 (re)transmission queue 2 1 2 3 4 retransmission number retransmission of ENHANCEMENT: green packet remove exponential backoff due to timeout LCN 2006, Tampa, FL, USA, November 2006 Network 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Enhancement: Fast Retransmit Bundling sender receiver ENHANCEMENT: piggyback all chunks in retransmission queue blue packet is NOT piggybacked when dupACKs (but would be if due to timeout) Network no no no no SACK SACK SACK SACK retransmission queue retransmission of green packet (chunks) due to dupACKs LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Enhancements Modified retransmission timer removal of exponential backoff minRTO = 200 ms (as in TCP) Modified retransmission bundling always allow aggressive bundling for fast retransmit Modified fast retransmit tested fast retransmit after 1 SACK Thin stream detection fewer packets in flight to trigger a fast retransmit added tracking of outstanding packets less than 4 in flight = thin stream LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Enhancement: Results (200 ms) Considerable reduction in average and maximum latencies Increase in number of fast retransmissions compared to timeouts Increase in number of retransmissions original SCTP Timeout Fast retransmit Total LCN 2006, Tampa, FL, USA, November 2006 reduced minRTO & fast retransmit modified timer restart no SACK delay 266 197 331 633 35 284 288 1 301 481 619 634 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen New lksctp versions & New Test Set Up New lksctp versions has been developed lksctp in 2.6.16 (2.5.72-0.7.1) only one retransmission due to fast retransmit, next timeout only 3 SACKs required for fast retransmits lksctp in 2.6.17 has no major changes for our scenario New tests 100 B packets RTTs: 0, 50, 100, 150, 200, 250 ms Packet inter-arrival times: 50, 100, 150, 200, 250 ms Dynamic thin stream detection WEB SCTP Many web-connections generating cross traffic (and thus losses) LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Results: New lksctp Still high average and worst case latencies LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Results: Fast Retransmission Modification Fast retransmit modification – 1 SACK Reduction in maximum and average latency As expected a large increase in fast retransmit An increase in spurious retransmissions LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Results: Removed Exponential Backoff Removed exponential backoff Reduction in maximum and average latency An increase in spurious retransmissions Retransmission aggressiveness does not really pose a congestion threat since the amount of data waiting to be sent is always less than the minimum transmission window LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Results: Reduced Minimum Time-out Reduced minRTO Faster timeouts Reduction in maximum and average latency An increase in spurious retransmissions LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Results: All Modifications Combined All thin stream modifications A further reduction in maximum and average latency As expected an increase in fast retransmit An increase in spurious retransmissions LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen Conclusions Based on SCTP description we expected (hoped for) reduced latencies compared to TCP Enhancements like reduced minRTO removal of exponential backoff removal of delayed SACKs … reduce latencies for thin streams The enhancements increase the number of spurious retransmissions, but maybe not important for thin streams!!?? LCN 2006, Tampa, FL, USA, November 2006 2006 Jon Pedersen, Carsten Griwodz & Pål Halvorsen