1 High performance Throughput Les Cottrell – SLAC Lecture # 5a presented at the 26 th International Nathiagali Summer College on Physics and Contemporary Needs, 25 th June – 14 th July, Nathiagali, Pakistan Partially funded by DOE/MICS Field Work Proposal on Internet End-to-end Performance Monitoring (IEPM), also supported by IUPAP

2 How to measure Selected about a dozen major collaborator sites in California, Colorado, Illinois, FR, CH, UK over last 9 months –Of interest to SLAC –Can get logon accounts Use iperf –Choose window size and # parallel streams –Run for 10 seconds together with ping (loaded) –Stop iperf, run ping (unloaded) for 10 seconds –Change window or number of streams & repeat Record streams, window, throughput (Mbits/s), loaded & unloaded ping responses

3 Default window size SLAC to CERN thruput vs windows & streams Hi-perf = big windows & multiple streams Improves ~ linearly with streams for small windows 8kB 16kB 32kB 100kB 1MB 64kB

4 E.g. thruput vs windows & streams ANL Colorado IN2P3, FR CERN, CH Caltech Window Mbits/s Streams I NFN, IT Mbits/s Daresbury, UK Mbits/s

5 Progress towards goal: 100 Mbytes/s Site-to-Site Focus on SLAC – Caltech over NTON; Using NTON wavelength division fibers up & down W. Coast US; Replaced Exemplar with 8*OC3 & Suns with Pentium IIIs & OC12 (622Mbps) SLAC Cisco with OC48 (2.4Gbps) and 2 × OC12; Caltech Juniper M160 & OC48 ~500 Mbits/s single stream achieved recently over OC12.

6 SC2000 WAN Challenge SC2000, Dallas to SLAC RTT ~ 48msec –SLAC/FNAL booth: Dell PowerEdge PIII 2 * 550MHz with 64bit PCI + Dell 850MHz both running Linux, each with GigE, connected to Cat 6009 with 2GigE bonded to Extreme SC2000 floor switch –NTON: OC48 to GSR to Cat 5500 Gig E to Sun E4500 4*460MHz and Sun E4500 6*336MHz Internet 2: 300 Mbits/s NTON 960Mbits/s Details: –www-iepm.slac.stanford.edu/monitoring/bulk/sc2k.htmlwww-iepm.slac.stanford.edu/monitoring/bulk/sc2k.html

7 Iperf throughput conclusions 1/2 Can saturate bottleneck links For a given iperf measurement, streams share throughput equally. For small window sizes throughput increases linearly with number of streams Predicted optimum window sizes can be large (> Mbyte) Need > 1 stream to get optimum performance Can get close to max thruput with small (<=32Mbyte) with sufficient (5-10) streams Improvements of 5 to 60 in thruput by using multiple streams & larger windows Loss not sensitive to throughput

8 Iperf thruput conclusions 2/2 For fixed streams*window product, streams are more effective than window size: There is an optimum number of streams above which performance flattens out See www-iepm.slac.stanford.edu/monitoring/bulk/www-iepm.slac.stanford.edu/monitoring/bulk/ 4.6Mbits/s864kBCaltech 1.7Mbits/s2256kBCaltech 26.8Mbits/s864kBCERN 9.45Mbits/2256kBCERN ThroughputStreamsWindowSite

9 Network Simulator (ns-2) From UCB, simulates network –Choice of stack (Reno, Tahoe, Vegas, SACK…) –RTT, bandwidth, flows, windows, queue lengths … Compare with measured results –Agrees well –Confirms observations (e.g. linear growth in throughput for small window sizes as increase number of flows)

10 Agreement of ns2 with observed

11 Ns-2 thruput & loss predict Indicates on unloaded link can get 70% of available bandwidth without causing noticeable packet loss Can get over 80-90% of available bandwidth Can overdrive: no extra throughput BUT extra loss 90%

12 Simulator benefits No traffic on network (nb throughput can use 90%) Can do what if experiments No need to install iperf servers or have accounts No need to configure host to allow large windows BUT –Need to estimate simulator parameters, e.g. RTT use ping or synack Bandwidth, use pchar, pipechar etc., moderately accurate AND its not the real thing –Need to validate vs. observed data –Need to simulate cross-traffic etc

13 Impact of cross-traffic on Iperf between SLAC & GSFC/ Maryland SCP HTTP bbftp iperf All TCP traffic Iperf port traffic To SLAC From SLAC

14 Impact on Others Make ping measurements with & without iperf loading –Loss loaded(unloaded) –RTT

15 Impact of applying QoS Defined 3 classes of service, application marked packets: –Scavenger service (1%), Best effort, & Priority service (30%) –Used DiffServ features in Cisco 7507 with DS3 link Appears to work as expected Measurements made by Dave Hartzell, of GreatPlains net, May 01

16 Improvements for major International BaBar sites Throughput improvements of 1 to 16 times in a year Links are being improved: ESnet, PHYnet, GARR, Janet, TEN-155 Improvements to come: IN2P3 => 155Mbps RAL => 622Mbps

17 Gigabit/second networking The start of a new era: –Very rapid progress towards 10Gbps networking in both the Local (LAN) and Wide area (WAN) networking environments are being made. –40Gbps is in sight on WANs, but what after? –The success of the LHC computing Grid critically depends on the availability of Gbps links between CERN and LHC regional centers. What does it mean? –In theory: 1GB file transferred in 11 seconds over a 1Gbps circuit (*) 1TB file transfer would still require 3 hours and 1PB file transfer would require 4 months –In practice: major transmission protocol issues will need to be addressed (*) according to the 75% empirical rule CERN

18 Very high speed file transfer (1) –High performance switched LAN assumed: requires time & money. –High performance WAN also assumed: also requires money but is becoming possible. very careful engineering mandatory. –Will remain very problematic especially over high bandwidth*delay paths: Might force the use Jumbo Frames because of interactions between TCP/IP and link error rates. –Could possibly conflict with strong security requirements CERN

19 CERN Very high speed file transfer (2) Following formula proposed by Matt Mathis/PSC (“The Macroscopic Behavior of the TCP Congestion Avoidance Algorithm”) to approximate the maximum TCP throughput under periodic packet loss: (MSS/RTT)*(1/sqrt(p)) where MSS is the maximum segment size, 1460 bytes, in practice,and “p” is the packet loss rate. Are TCP's "congestion avoidance" algorithms compatible with high speed, long distance networks. –The "cut transmit rate in half on single packet loss and then increase the rate additively (1 MSS by RTT)" algorithm may simply not work. –New TCP/IP adaptations may be needed in order to better cope with “lfn”, e.g. TCP Vegas

20 Acceptable link error rates CERN

21 Very high speed file transfer (tentative conclusions) Tcp/ip fairness only exist between similar flows, i.e. similar duration, similar RTTs. Tcp/ip congestion avoidance algorithms need to be revisited (e.g. Vegas rather than Reno/NewReno) –faster recovery after loss, selective acknowledgment. Current ways of circumventing the problem, e.g. –Multi-stream & parallel socket just bandages or the practical solution to the problem? Web100, a 3MUSD NSF project, might help enormously! better TCP/IP instrumentation (MIB), will allow read/write to internal TCP parameters self-tuning tools for measuring performance improved FTP implementation applications can tune stack Non-Tcp/ip based transport solution, use of Forward Error Corrections (FEC), Early Congestion Notifications (ECN) rather than active queue management techniques (RED/WRED)? CERN

22 Optimizing streams Choose # streams to optimize throughput/impact –Measure RTT from Web100 –App controls # streams

23 WAN thruput conclusions High FTP performance across WAN links is possible –Even with 20-30Mbps bottleneck can do > 100Gbytes/day OS must support big windows selectable by application Need multiple parallel streams Loss is important in particular interval between losses Compression looks promising, but needs cpu power Can get close to max thruput with small (<=32Mbyte) with sufficient (5-10) streams Improvements of 5 to 60 in thruput by using multiple streams & larger windows Impacts others users, need Less than Best Effort QoS service

24 More Information This talk: – IEPM/PingER home site –www-iepm.slac.stanford.edu/www-iepm.slac.stanford.edu/ Transfer tools: – TCP Tuning: –

25 High Speed Bulk Throughput Driven by: –Data intensive science, e.g. data grids –HENP data rates, e.g. BaBar 300TB/year, collection doubling yearly, i.e. PBytes in couple of years –Data rate from experiment ~ 20MBytes/s ~ 200GBytes/d –Multiple regional computer centers (e.g. Lyon-FR, RAL-UK, INFN-IT, LBNL-CA, LLNL-CA, Caltech-CA) need copies of data –Boeing 747 high throughput, BUT poor latency (~ 2 weeks) & very people intensive So need high-speed networks and ability to utilize –High speed today = few hundred GBytes/day (100GB/d ~ 10Mbits/s) Data vol Moore’s law