Network-aware OS DOE/MICS Project Final Review September 16, 2004 Tom Dunigan Matt Mathis Brian Tierney ORNL : Florence Fowler, Steven Carter, Nagi Rao, Bill Wing PSC : Raghu Reddy, John Heffner, Janet Brown LBNL : Jason Lee, Martin Stouffer

U.S. Department of Energy Office of Science LBNL/ORNL/PSC Roadmap Motivation & Background Net100 project components –Web100 –network probes & sensors –protocol analysis and tuning Results –TCP tuning daemon –Tuning experiments Project contributions rg DOE-funded project (Office of Science) $2.6M, 3 yrs beginning 9/01 LBNL, ORNL, PSC, NCAR Net100 project objectives: (network-aware operating systems) measure, understand, and improve end-to-end network/application performance tune network protocols and applications (grid and bulk transfer) emphasis: TCP bulk transfer over high delay/bandwidth nets

U.S. Department of Energy Office of Science LBNL/ORNL/PSC Motivation Poor network application performance –High bandwidth paths, but app’s slow – Is it application? OS? network? … Yes –Often need a network “wizard” Changing : bandwidths –9.6 Kbs… 1.5 Mbs..45 …100…1000…? Gbs Unchanging: TCP –speed of light (RTT) –packet size (MSS/MTU) still 1500 bytes –TCP congestion control TCP is lossy by design ! –2x overshoot at startup, sawtooth –Recovery proportional to MSS/RTT 2 –recovery after a loss can be very slow on today’s high delay/bandwidth links -- unacceptable on tomorrow’s links: 10 Gbs cross country: recovery time > 1 hr.!! Linear recovery at 0.5 Mb/s! Instantaneous bandwidth Average bandwidth Early startup losses ORNL to NERSC ftp 8 Mbs GigE/OC12 (600 Mbs) 80ms RTT 40 seconds

U.S. Department of Energy Office of Science LBNL/ORNL/PSC TCP 101 adaptable and fair flow-controlled by sender/receiver buffer sizes self-clocking with positive ACK’s of in-sequence data sensitive to packet size (MTU) and RTT slow start packet per each packet ACK’d (exponential) congestion window ( cwnd )-- max packets that can be in flight packet loss: 3 dup ACKs or timeout (AIMD) –cut cwnd in half (Multiplicative Decrease) – add 1 packet to cwnd per RTT (Additive Increase) Workarounds: – parallel streams – non-TCP (UDP) applications – Net100 ( no changes to applications )

U.S. Department of Energy Office of Science LBNL/ORNL/PSC Net100 components Web100 Linux kernel (NSF) –instrumented TCP stack (IETF MIB draft) Path characterization –Network Tuning and Analysis Framework (NTAF) –both active and passive measurement tools –data base of measurements TCP protocol analysis and tuning –simulation/emulation ns TCP-over-UDP ( atou ) NISTNet –kernel tuning extensions –tuning daemon (WAD)

U.S. Department of Energy Office of Science LBNL/ORNL/PSC Web100 NSF funded (PSC/NCAR/NCSA) web100.org Modified Linux kernel – instrumented kernel to read/set TCP variables for a specific flow –readable: RTT, counts (bytes, pkts, retransmits,dups), state (SACKs, windowscale, cwnd, ssthresh) –settable: buffer sizes –100+ TCP variables (IETF MIB) ( /proc/web100/) GUI to display/modify a flow’s TCP variables, real-time API for network-aware applications or tuning daemon Net100 extensions: –additional tuning variables and algorithms –event notification for a tuning daemon –Java bandwidth tester cruise.ornl.gov:7123

U.S. Department of Energy Office of Science LBNL/ORNL/PSC Network Tool Analysis Framework (NTAF) Configure and launch network tools –measure bandwidth/latency ( iperf, pchar, pipechar ) –augment tools to report Web100 data Collect and transform tool results –use Netlogger to transform common format Save results for short-term auto-tuning and archive for later analysis –compare predicted to actual performance –measure effectiveness of tools and auto-tuning –provide data that can be used to predict future performance –invaluable for comparing tools ( pathload/pchar/netest ) Net100 hosts at : LBNL,ORNL,PSC,NCAR NERSC, SLAC, UT, CERN, Amsterdam,ANL

U.S. Department of Energy Office of Science LBNL/ORNL/PSC TCP flow visualization - Web interface for data archive and visualization

U.S. Department of Energy Office of Science LBNL/ORNL/PSC TCP tuning “enable” high speed –need buffer = bandwidth*RTT - autotune ORNL/NERSC (80 ms, OC12) need 6 MB –faster slow-start avoid losses –modified slow-start –reduce bursts –anticipate loss (ECN,Vegas?) –reorder threshold speed recovery –bigger MTU or “virtual MSS” –modified AIMD (0.5,1) (Floyd, Kelly) –delayed ACKs, initial window, slow-start increment avoid congestion collapse, be fair (?) … intranets, QoS Net100: ns simulation, NISTNet emulation, “almost TCP over UDP” ( atou ), WAD/Internet ns simulation: 500 mbs link, 80 ms RTT Packet loss early in slow start. Standard TCP with del ACK takes 10 minutes to recover!

U.S. Department of Energy Office of Science LBNL/ORNL/PSC TCP Tuning Daemon Work-around Daemon (WAD) –tune unknowing sender/receiver at startup and/or during flow –Web100 kernel extensions pre-set windowscale to allow dynamic tuning uses netlink to alert daemon of socket open/close (or poll) besides existing Web100 buffer tuning, new tuning parameters and algorithms knobs to disable Linux 2.4 caching, burst mgt., and sendstall –config file with static tuning data mode specifies dynamic tuning (AIMD options, NTAF buffer size, concurrent streams) –daemon periodically polls NTAF for fresh tuning data –can do out-of-kernel tuning (e.g., Floyd) –written in C (also Python version) WAD config file [bob] src_addr: src_port: 0 dst_addr: dst_port: 0 mode: 1 sndbuf: rcvbuf: wadai: 6 wadmd: 0.3 maxssth: 100 divide: 1 reorder: 9 sendstall: 0 delack: 0 floyd: 1 kellyai: 0

U.S. Department of Energy Office of Science LBNL/ORNL/PSC Experimental results Evaluating the tuning daemon in the wild –emphasis: bulk transfers over high delay/bandwidth nets (Internet2, ESnet) –tests over: 10GigE/OC192,OC48, OC12, OC3, ATM/VBR, GigE+jumboframe,FDDI,100/10T,cable, ISDN,wireless (802.11b),dialup –tests over NISTNet testbed (speed, loss, delay) Various TCP tuning options –buffer tuning (static, auto, and dynamic/NTAF) –AIMD mods (including Floyd, Kelly, Vegas, static, virtual MSS) –slow-start mods –parallel streams vs single tuned vs UDP transports NISTNet host

U.S. Department of Energy Office of Science LBNL/ORNL/PSC Buffer tuning Classic buffer tuning network-challenged app. gets 10 Mbs same app., WAD/NTAF tuned buffer gets 143 Mbs Autotuning buffers ( kernel ) Linux 2.4, Feng’s Dynamic Right Sizing Net100 autotuning receiver estimates RTT receiver advertises window 2 times data recv’d in RTT buffer size grows dynamically to 2x bandwidth*RTT separate application buffers from kernel buffers ORNL to PSC, OC192, 30 ms RTT ORNL to PSC, OC12, 80ms RTT

U.S. Department of Energy Office of Science LBNL/ORNL/PSC Speeding recovery Amsterdam-Chicago GigE via 10GigE, 100 ms RTT UDP burst Selectable TCP AIMD algorithms : Floyd HS TCP : as cwnd grows increase AI and decrease MD, do the reverse when cwnd shrinks Kelly scalable TCP : use MD of 1/8 instead of 1/2 and add % of cwnd (e.g. 1%) each RTT Virtual MSS tune TCP’s additive increase (WAD_AI) add k segments per RTT during recovery k =6 like GigE jumbo frame, but: interrupt rate not reduced doesn’t do k segments for initial window

U.S. Department of Energy Office of Science LBNL/ORNL/PSC WAD tuning Modified slow-start and AI often losses in slow-start WAD tuned Floyd slow-start and fixed AI (6) WAD-tuned AIMD and slow-start parallel streams AIMD (1/(2k),k) exploit TCP’s fairness WAD-tuned single stream (0.125,4) “ “ + Floyd slow-start ORNL to NERSC, OC12, 80 ms RTT ORNL to CERN, OC12, 150ms RTT

U.S. Department of Energy Office of Science LBNL/ORNL/PSC Workaround: parallel streams Takes advantage of TCP’s fairness Faster startup, k buffers faster recovery –often only 1 stream loses a packet –MD: 1/(2k) rather than 1/2 –AI: k times faster linear phase BUT –requires rewrite of applications –how many streams? Buffer size? GridFTP, bbftp, psocket lib Alice and Bob sharing Clever Alice -- 3 streams Bad girl...

U.S. Department of Energy Office of Science LBNL/ORNL/PSC GridFTP tuning Can tuned single stream compete with parallel streams? Mostly not with “equivalence” tuning, but sometimes…. Parallel streams have slow-start advantage. WAD can divide buffer among concurrent flows—fairer/faster? Tests inconclusive Testing on real Internet is problematic. Is there a “congestion metric”? Per unit of time? Flow Mbs congestion re-xmits untuned tuned parallel untuned tuned parallel Data/plots from Web100 tracer Buffers: 64K I/O, 4MB TCP

U.S. Department of Energy Office of Science LBNL/ORNL/PSC Recent Net100 research –more user-friendly WAD, WAD-lite No NTAF, bandwidth test thread –invited to submit Web100/Net100 mods to Linux 2.6 –port to Cray X1 Linux network front-end added Net100 kernel, 4x improvement in wide-area TCP! –port to SGI Altix –TCP Vegas Vegas avoids loss (if RTT increasing, Vegas backs off) can be configured to compete with standard TCP (Feng) CalTech’s FAST (adjusts alpha dynamically) –comparison with other “work arounds” parallel streams non-TCP (SABUL, FOBS, TSUNAMI, RBUDP, SCTP) –additional accelerants slow-start initial/increment reorder resiliance delayed ACKs

U.S. Department of Energy Office of Science LBNL/ORNL/PSC TCP tuning for other OS’s Reorder threshold seeing more out of order packets - future: multipath or bonded NICs WAD tune a bigger reorder threshold for path 40x improvement! Linux 2.4 does a good job already adjusts and caches reorder threshold “undo” congestion avoidance UDP transports don’t handle re-ordering well Delayed ACKs WAD could turn off delayed ACKs 2x improvement in recovery rate and slow-start Linux 2.4 already turns off delayed ACKs for initial slow-start ns simulation: 500 mbs link, 80 ms RTT Packet loss early in slow-start. Standard TCP with del ACK takes 10 minutes to recover! NOTE aggressive static AIMD (Floyd pre-tune) LBL to ORNL (using our TCP-over-UDP) : dup3 case had 289 retransmits, but all were unneeded!

U.S. Department of Energy Office of Science LBNL/ORNL/PSC Interactions Scientific applications –SciDAC supernova and global climate –Data grids (CERN, SLAC) Middleware –Globus/gridFTP –HSI/HPSS Network measurement –Internet2 end-to-end –Pinger (Cottrell) –Claffy/Dovrolis pathload –netest (Guojun) –SCNM Protocol research –Dynamic Right-Sizing (Feng) –HS TCP (Floyd) –Scalable TCP (Kelly) –TCP Vegas (Feng, Low) –Tsunami/SABUL/FOBS/RBUDP –parallel streams (Hacker) OS vendors –Linux –IBM AIX/Linux –Cray X1 –SGI Altix Talks/papers/software/

U.S. Department of Energy Office of Science LBNL/ORNL/PSC Insights Parallel streams are quite effective –No kernel mods, but need new app’s –Bypass system buffer limits –Faster slow-start and recovery, and still TCP-like Rate-based UDP is effective –No kernel mods, but need new app’s –Sensitive to re-ordering –Many duplicate packets –Does software-based rate control in the application layer scale? WAD and WAD-lite: nice for experimenting or QoS, hard for user –Configure auto-tuning and Floyd’s HS TCP Vote for bigger MTUs

U.S. Department of Energy Office of Science LBNL/ORNL/PSC Summary Novel approaches –non-invasive dynamic/auto tuning of legacy applications –out-of-kernel tuning –using TCP to tune TCP –tuning on a per flow/destination based on recent path metrics or policy (QoS) Effective evaluation framework –protocol analysis and tuning –network/application/OS debugging –path characterization tools, archive, and visualization tools Performance improvements –WAD tuned : buffers  10x AIMD  2x to 10x delayed ACK  2x slowstart  3x reorder  40x Timely -- needed for science on today’s and tomorrow’s networks