Evaluation of Advanced TCP stacks on Fast Long-Distance production Networks

1. Evaluation of Advanced TCP stacks on Fast Long-Distance production Networks Prepared by Les Cottrell & Hadrien Bullot, Richard Hughes-Jones EPFL, SLAC and Manchester University for the Protocols for Fast Long Distance Networks, ANL February, 2004 www.slac.stanford.edu/grp/scs/net/talk03/pfld-feb04d.ppt Partially funded by DOE/MICS Field Work Proposal on Internet End-to-end Performance Monitoring (IEPM), also supported by IUPAPand UK e-Science via PPARC

2. Project goals • Test new advanced TCP stacks, see how they perform on short and long-distance real production WAN links • Compare & contrast: ease of configuration, throughput, convergence, fairness, stability etc. • For different RTTs, windows, txqueuelen • Recommend “optimum” stacks for data intensive science: (BaBar) transfers using bbftp, bbcp, GridFTP • Validate simulator & emulator findings & provide feedback

3. Protocol selection • Focus on TCP only • No Rate based transport protocols (e.g. SABUL, UDT, RBUDP) at the moment • No iSCSI or FC over IP • Sender mods only, HENP model is few big senders, lots of smaller receivers • Simplifies deployment, only a few hosts at a few sending sites • No DRS • Runs made on production networks so: • No router mods (XCP/ECN), no jumbos,

4. Protocols Evaluated • Linux 2.4 New Reno with SACK: single (Reno) and parallel streams (P-TCP) • Scalable TCP (S-TCP) • Fast TCP • HighSpeed TCP (HS-TCP) • HighSpeed TCP Low Priority (HSTCP-LP) • Binary Increase Control TCP (Bic-TCP) • Hamilton TCP (H-TCP)

5. Reno single stream • Low performance on fast long distance paths • AIMD (add a=1 pkt to cwnd / RTT, decrease cwnd by factor b=0.5 in congestion) 700 SLAC to Florida Throughput Mbps Reno RTT ms RTT (~70ms) 0 1200 s

6. Measurements Over a thousand 20 minute measurements or 300 hours 600 Mbps capacity Utilization of SLAC ESnet link Sep-Nov ‘03 • 20 minute tests, long enough to see stable patterns • Iperf reports incremental and cumulative throughputs at 5 second intervals • Ping interval about 100ms Ping traffic goes to TCPr when also running cross-traffic Otherwise goes to Xr SLAC UDP or TCP cross-traffic Xs Xr bottleneck TCP Remote site TCPs TCPr ping ICMP/ping traffic

7. Networks • 3 main network paths • Short distance: SLAC-Caltech (RTT~10ms) • Middle distance: U. Florida (UFl) & DataTAG Chicago( RTT~70ms) • Long distance: CERN & University of Manchester (RTT ~ 170ms) • Tests during nights and weekends to avoid unacceptable impacts on production traffic

8. Windows • Set large maximum windows (typically 32MB) on all hosts • Used 3 different windows with iperf: • Small window size, factor 2-4 below optimal • Roughly optimal window size (~BDP) • Oversized window

9. RTT • Only P-TCP appears to dramatically affect the RTT • E.g. increases by RTT by 200ms (factor 20 for short distances) • Implication: P-TCP would impact apps. like Voice/IP 700 600 700 600 SLAC-Caltech P-TCP 16 stream SLAC-Caltech FAST TCP 1 stream RTT (ms) RTT (ms) Throughput (Mbps) RTT Throughput (Mbps) RTT 0 0 1200 1200 Time (secs) Time (secs)

10. txqueuelen • Regulates the size of the queue between the IP layer and the Ethernet layer • May increase the throughput if we find optimal values • But may increase duplicate ACKs (Y. T Li) • All stacks except S-TCP use txqueuelen=100 as default • S-TCP uses txqueuelen=2000 by default • Tests showed these were reasonable choices

11. Throughput (Mbps) Windows too small (worse for longer distance) Poor performance Reasonable performance Better performance Best performance Reno with 1 stream has problems on Medium distance link (70ms) Window size ?

12. Throughput Avg throughput for optimal & large window sizes from SLAC to CalTech, UFl & Manchester Stack more important for long RTTs Single stream Reno & HSTCP-LP poorer on large RTTs

13. Stability • Definition: standard deviation normalized by the average throughput • At short RTT (10ms) stability is usually good (<=12%) • At medium RTT (70ms) P-TCP, Scalable & Bic-TCP appear more stable than the other protocols SLAC-UFl Scalable 8MB window, txq=2000, 380Mbps SLAC-UFl FAST 8MB window, txq=100, 350Mbps Stability ~ 0.098 Stability ~ 0.28

14. Sinusoidal UDP • UDP does not back off in face of congestion, it has a “stiff” behavior • We modified iperf to allow it to create UDP traffic with a sinusoidal time behavior, following an idea from Tom Hacker • See how TCP responds to varying cross-traffic • Used 2 periods of 30 and 60 seconds and amplitude varying from 20 to 80 Mbps • Sent from 2nd sending host to 2nd receiving host while sending TCP from 1st sending host to 1st receiving host • As long as the window size was large enough all protocols converged quickly and maintain a roughly constant aggregate throughput • Especially for P-TCP & Bic-TCP

15. TCP Stability against UDP • Stability better at short distances • P-TCP & Bic more stable 700 600 SLAC-UFl H-TCP Stability~0.27, 276Mbps SLAC-UFl Bic-TCP Stability~0.11, 355Mbps Aggregate RTT (ms) Throughput (Mbps) TCP RTT UDP 0 1200 Time (secs)

16. Stability Stability from SLAC to Caltech, U Florida & Manchester No UDP +UDP Stability & distanceShort RTT is more stable Little difference between periodicity of UDP (30 & 60 secs) HSTCP-LP & FAST have larger stability indices (less stability)

17. Cross TCP Traffic • Important to understand how fair a protocol is • For one protocol competing against the same protocol (intra-protocol) we define the fairness for a single bottleneck as: • All protocols have good intra-protocol Fairness (F>0.98) • Except HS-TCP (F<0.94) when the window size > optimal 700 600 700 600 SLAC-Caltech 70 ms Fast-TCP (F~0.997) SLAC-Florida 10ms HS-TCP (F~0.935) Aggregate Aggregate RTT (ms) Throughput (Mbps) Throughput (Mbps) RTT (ms) TCPs TCP RTT RTT 1200 Time (secs) Time (secs) 1200

18. Fairness (F) • Most have good intra-protocol fairness (diagonal elements), except HS-TCP • Worse for larger RTT (Caltech F~0.999+-0.004, U Florida F~0.995+-0.14, Manchester F~0.95+-0.05) • Inter protocol Bic & H appear more fair against others • Worst fairness are: P-TCP, S-TCP, Fast, HSTCP-LP (backoff early) • But cannot tell who is aggressive and who is timid

19. Inter protocol Fairness • For inter-protocol fairness we introduce the asymmetry between the two throughputs: • Where x1and x2 are the throughput averages of TCP stack 1 competing with TCP stack 2 +ve TCP1 agressive Avg. Asymmetry vs all stacks Avg. Asymmetry vs all stacks Reno 16 v. aggressive at short RTT, Reno 16 & Scalable aggressive at medium distance on medium RTT: HS-TCP timidHSTCP-LP very timid

20. Inter Fairness – UFl (A) A=(xm-xc)/(xm+xc) Aggressive Fair Timid Diagonal = 0 by definition Symmetric off diagonal Down how does X traffic behave Scalable & Reno 16 streams are aggressive HS LP is very timid Fast more aggressive than HS & H HS is timid

21. Reverse Traffic • Cause queuing on reverse path by using P-TCP 16 streams • ACKs are lost or come back in bursts (compressed ACKs) • Fast TCP throughput is 4 to 8 times less than the other TCPs. 700 600 600 700 SLAC-Florida Fast TCP SLAC-Florida Bic-TCP Reverse traffic Push up rtt Reverse traffic RTT (ms) Throughput (Mbps) Throughput (Mbps) RTT (ms) RTT TCP RTT 1200 1200 Time (secs) Time (secs)

22. SC2003 Phoenix "Bandwidth Lust"

23. SC2003 Bandwidth Challenge:Network MAP Paix UvA Starlight

24. 10 GigEthernet TCP Flows • Three Server systems with Intel 10 GigEthernet NICs • Used the DataTAG altAIMD stack and 9000 byte MTU • Streams From SLAC/FNAL booth in Phoenix to: • Pal Alto PAIX • Min rtt 17 ms, window 30 MB • Shared with FAST TCP fromCaltech booth • 4.37 Gbit hstcp I=0.5% • then 2.87 Gbit I=16% • Fall corresponds to 10 Gbit on link • 3.3Gbit Scalable I=8% • Tested 2 flows, sum 1.9Gbit I=39% • Used Abilene to Chicago • Chicago Starlight Itanium • Min rtt 65 ms , window 60 MB • Phoenix CPU 2.2 GHz • 3.1 Gbit hstcp I=1.6% • Amsterdam SARA Itanium • Min rtt 175 ms , window 200 MB • Phoenix CPU 3.06 GHz • 4.35 Gbit hstcp I=6.9% • Very Stable

25. Preliminary Conclusions • Advanced stacks behave like TCP-Reno single stream on short distances for up to Gbits/s paths, especially if window size limited • TCP Reno single stream has low performance and is unstable on long distances • P-TCP is very aggressive and impacts the RTT badly • HSTCP-LP is too gentle, this can be important for providing scavenger service without router modifications. By design it backs off quickly, otherwise performs well • Fast TCP is very handicapped by reverse traffic • S-TCP is very aggressive on long distances • HS-TCP is very gentle, like H-TCP has lower throughput than other protocols • Bic-TCP performs very well in almost all cases

26. Current/Future work • Work with Caltech to correlate with simulation • Compare with other people’s measurements • Test Westwood+, LTCP • Testing UDP rate based transports: UDT • Looks good, BUT 4*CPU utilization of TCP • Try on 10Gbps links • More tests with multiple streams • Use with production applications: IN2P3

27. More Information • TCP Stacks Evaluation: • www-iepm.slac.stanford.edu/bw/tcp-eval/ • www.slac.stanford.edu/grp/scs/net/papers/hadrien/report/report2.pdf

28. Extra Slides

29. P-TCP • TCP Reno with 16 streams • Parallel streams heavily used in HENP & elsewhere to achieve needed performance, so it is today’s de facto baseline • However, hard to optimize both the window size AND number of streams since optimal values can vary due to network capacity, routes or utilization changes SLAC Caltech May – Nov ’03, note weekend/weekday changes & upgrades

30. S-TCP • Uses exponential increase everywhere (in slow start and congestion avoidance) • Multiplicative decrease factor b = 0.125 • Introduced by Tom Kelly of Cambridge

31. Fast TCP • Based on TCP Vegas • Uses both queuing delay and packet losses as congestion measures • Developed at Caltech by Steven Low and collaborators

32. HS-TCP • Behaves like Reno for small values of cwnd • Above a chosen value of cwnd (default 38) a more aggressive function is used • Uses a table to indicate by how much to increase cwnd when an ACK is received • Introduced by Sally Floyd

33. HSTCP-LP • Mixture of HS-TCP with TCP-LP (Low Priority) • Backs off early in face of congestion by looking at RTT • Idea is to give scavengers service without router modifications • From Rice University

34. Bic-TCP • Combine: • An additive increase used for large cwnd • A binary search increase used for small cwnd • Developed Injong Rhee at NC State University

35. H-TCP • Similar to HS-TCP in switching to aggressive mode after threshold • Uses an heterogeneous AIMD algorithm • Developed at Hamilton U Ireland

36. Throughput • With optimal window all stacks within ~20% of one another, except Reno 1 stream on medium and long distances • P-TCP & S-TCP get best throughput

37. Inter Fair Caltech A= (x1-x2) (x1+x2) Aggressive Fair Timid Less inter protocol differences than for UFL (10ms vs 70ms) Everyone timid in presence of Reno 16 streams (even Scalable)

38. Stability Avg. Stability from SLAC to Caltech, UFl & Manchester for optimal & large windows with no competing UDP streams Avg. Stability from SLAC to Caltech, UFl & Manchester for optimal & large windows with competing UDP streams HSTCP-LP & FAST TCP less stable Bic-TCP, S-TCP, Reno-16, H-TCP more stable