GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 1 End-2-End Network Monitoring What do we do ? What do we use it for? Richard Hughes-Jones Many people are involved:

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 2 Local Network Monitoring Store & Analysis of Data (Access) Access to current and historic data and metrics via the Web, i.e. WP7 NM Pages, access to metric forecasts Backend LDAP script to fetch metrics Monitor process to push metrics local LDAP Server Grid Application access via LDAP Schema to - monitoring metrics; - location of monitoring data. PingER (RIPE TTB) iperf UDPmon rTPL NWS etc LDAP Schema Grid Apps GridFTP DataGRID WP7: Network Monitoring Architecture for Grid Sites Robin Tasker

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 3 WP7 Network Monitoring Components PingNetmonUDPmoniPerfRipe Cron script plot Table LDAP raw control Cron script control Cron script plot Table LDAP rawplot Table LDAP raw WEB DisplayAnalysisGrid BrokersPredictions Web I/f Scheduler Tool Clients LDAP raw LDAP raw

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 4  Grid network monitoring architecture uses LDAP & R-GMA - DataGrid WP7  Central MySQL archive hosting all network metrics and GridFTP logging  Probe Coordination Protocol deployed, scheduling tests  MapCentre also provides site & node Fabric health checks WP7 MapCentre: Grid Monitoring & Visualisation Franck Bonnassieux CNRS Lyon

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 5 CERN – RAL UDP CERN – IN2P3 UDP WP7 MapCentre: Grid Monitoring & Visualisation CERN – RAL TCP CERN – IN2P3 TCP

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 6 UK e-Science: Network Monitoring  Technology Transfer  DataGrid WP7 M/c  UK e-Science DL  DataGrid WP7 M/c  Architecture

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 7 UK e-Science: Network Problem Solving 24 Jan to 4 Feb 04 TCP iperf RAL to HEP Only 2 sites >80 Mbit/s RAL -> DL Mbit/s 24 Jan to 4 Feb 04 TCP iperf DL to HEP DL -> RAL ~80 Mbit/s

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 8  UDP/IP packets sent between end systems  Latency  Round trip times using Request-Response UDP frames  Latency as a function of frame size Slope s given by: Mem-mem copy(s) + pci + Gig Ethernet + pci + mem-mem copy(s) Intercept indicates processing times + HW latencies  Histograms of ‘singleton’ measurements  UDP Throughput  Send a controlled stream of UDP frames spaced at regular intervals  Vary the frame size and the frame transmit spacing & measure: The time of first and last frames received The number packets received, lost, & out of order Histogram inter-packet spacing received packets Packet loss pattern 1-way delay CPU load Number of interrupts Tools: UDPmon – Latency & Throughput n bytes Number of packets Wait time time 

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 9 UDPmon: Example 1 Gigabit NIC Intel pro/1000 Latency Throughput Bus Activity  Motherboard: Supermicro P4DP6  Chipset: E7500 (Plumas)  CPU: Dual Xeon 2 2GHz with 512k L2 cache  Mem bus 400 MHz PCI-X 64 bit 66 MHz  HP Linux Kernel SMP  MTU 1500 bytes  Intel PRO/1000 XT Receive Transfer Send Transfer

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 10 Tools: Trace-Rate Hop by hop measurements  A method to measure the hop-by-hop capacity, delay, and loss up to the path bottleneck  Not intrusive  Operates in a high-performance environment  Does not need cooperation of the destination  Based on Packet Pair Method  Send sets of b2b packets with increasing time to live  For each set filter “noise” from rtt  Calculate spacing – hence bottleneck BW  Robust regarding the presence of invisible nodes Effect of the bottleneck on a packet pair. L is a packet size C is the capacity Examples of parameters that are iteratively analysed to extract the capacity mode

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 11 Tools: Trace-Rate Some Results  Capacity measurements as function of load in Mbit/s from tests on the DataTAG Link:  Comparison of the number of packets required  Validated by simulations in NS-2  Linux implementations, working in a high-performance environment  Research report:  Research Paper: ICC2004 : International Conference on Communications, Paris, France, June IEEE Communication Society.

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 12 Network Monitoring as a Tool to study:  Protocol Behaviour  Network Performance  Application Performance  Tools include:  web100  tcpdump  Output from the test tool: UDPmon, iperf, …  Output from the application Gridftp, bbcp, apache

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 13 Protocol Performance: RDUDP  Monitoring from Data Moving Application & Network Test Program  DataTAG WP3 work  Test Setup:  Path: Ams-Chi-Ams Force10 loopback  Moving data from DAS-2 cluster with RUDP – UDP based Transport  Apply 11*11 TCP background streams from iperf  Conclusions  RUDP performs well  It does Back off and share BW  Rapidly expands when BW free Hans Blom

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 14 Performance of the GÉANT Core Network  Test Setup:  Supermicro PC in: London & Amsterdam GÉANT PoP  Smartbits in: London & Frankfurt GÉANT PoP  Long link : UK-SE-DE2-IT-CH-FR-BE-NL  Short Link : UK-FR-BE-NL  Network Quality Of Service  LBE, IP Premium  High-Throughput Transfers  Standard and advanced TCP stacks  Packet re-ordering effects  Jitter for IPP and BE flows under load Flow: BE BG:60+40% BE+LBE Flow: IPP BG:60+40% BE+LBE Flow: IPP none

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 15 Tests GÉANT Core: Packet re-ordering Effect of LBE background Amsterdam-London BE Test flow Packets at 10 µs – line speed 10,000 sent Packet Loss ~ 0.1% Re-order Distributions:

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 16 Application Throughput + Web100  2Gbyte file transferred RAID0 disks  Web100 output every 10 ms  Gridftp  See alternate 600/800 Mbit and zero MB - NG  Apachie web server + curl-based client  See steady 720 Mbit

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester byte Packets man -> JIVE FWHM 22 µs (B2B 3 µs ) VLBI Project: Throughput Jitter 1-way Delay Loss 1-way Delay – note the packet loss (points with zero 1 –way delay) 1472 byte Packets Manchester -> Dwingeloo JIVE Packets Loss distribution Prob. Density Function: P(t) = λ e-λt Mean λ = 2360 / s [426 µs]

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 18 Passive Monitoring  Time-series data from Routers and Switches  Immediate but usually historical- MRTG  Usually derived from SNMP  Miss-configured / infected / misbehaving End Systems (or Users?)  Note Data Protection Laws & confidentiality  Site MAN and Back-bone topology & load  Help to user/sysadmin to isolate problem – eg low TCP transfer  Essential for Proof of Concept tests or Protocol testing  Trends used for capacity planning  Control of P2P traffic

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 19 Users: The Campus & the MAN [1]  NNW – to – SJ4 Access 2.5 Gbit PoS Hits 1 Gbit 50 %  Man – NNW Access 2 * 1 Gbit Ethernet Pete White Pat Myers

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 20 Users: The Campus & the MAN [2]  LMN to site 1 Access 1 Gbit Ethernet  LMN to site 2 Access 1 Gbit Ethernet  Message:  Not a complaint  Continue to work with your network group  Understand the traffic levels  Understand the Network Topology

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 21 VLBI Traffic Flows  Manchester – NetNorthWest - SuperJANET Access links  Two 1 Gbit/s  Access links: SJ4 to GÉANT GÉANT to SurfNet Only testing – Could be worse!

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 22  Network Measurement Working Group  “A Hierarchy of Network Performance Characteristics for Grid Applications and Services”  Document defines terms & relations:  Network characteristics  Measurement methodologies  Observation  Discusses Nodes & Paths  For each Characteristic  Defines the meaning  Attributes that SHOULD be included  Issues to consider when making an observation  Status:  Originally submitted to GFSG as Community Practice Document draft-ggf-nmwg-hierarchy-00.pdf Jul 2003  Revised to Proposed Recommendation 7 Jan 04  Now in 60 day Public comment from 28 Jan 04 – 18 days to go. GGF: Hierarchy Characteristics Document

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 23  Request Schema:  Ask for results / ask to make test  Schema Requirements Document made Use DAMED style names e.g. path.delay.oneWay Send: Char. Time, Subject = node | path Methodology, Stats  Response Schema:  Interpret results  Includes Observation environment  Much work in progress  Common components  Drafts almost done  2 (3) proof-of-concept implementations  2 implementations using XML-RPC by Internet2 SLAC  Implementation in progress using Document /Literal by DL & UCL Network Monitoring Service XML test request XML tests results GGF: Schemata for Network Measurements

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 24 So What do we Use Monitoring for: A Summary  End2End Time Series  Throughput UDP/TCP  Rtt  Packet loss  Passive Monitoring  Routers Switches SNMP MRTG  Historical MRTG  Packet/Protocol Dynamics  tcpdump  web100  Output from Application tools  Detect or X-check problem reports  Isolate / determine a performance issue  Capacity planning  Publication of data: network “cost” for middleware  RBs for optimized matchmaking  WP2 Replica Manager  Capacity planning  SLA verification  Isolate / determine throughput bottleneck – work with real user problems  Test conditions for Protocol/HW investigations  Protocol performance / development  Hardware performance / development  Application analysis  Input to middleware – eg gridftp throughput  Isolate / determine a (user) performance issue  Hardware / protocol investigations

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 25 More Information Some URLs  DataGrid WP7 Mapcenter: &  UK e-science monitoring:  MB-NG project web site:  DataTAG project web site:  UDPmon / TCPmon kit + writeup:  Motherboard and NIC Tests:  IEPM-BW site:

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 26

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 27  Network Monitoring to Grid Sites  Network Tools Developed  Using Network Monitoring as a Study Tool  Applications & Network Monitoring – real users  Passive Monitoring  Standards – Links to GGF

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 28 Data Flow: SuperMicro 370DLE: SysKonnect Motherboard: SuperMicro 370DLE Chipset: ServerWorks III LE Chipset CPU: PIII 800 MHz PCI:64 bit 66 MHz RedHat 7.1 Kernel bytes sent Wait 100 us ~8 us for send or receive Stack & Application overhead ~ 10 us / node Send PCI Receive PCI ~36 us Send Transfer Send CSR setup Receive Transfer Packet on Ethernet Fibre

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester GigEthernet: Throughput  1500 byte MTU gives ~ 2 Gbit/s  Used byte MTU max user length  DataTAG Supermicro PCs  Dual 2.2 GHz Xeon CPU FSB 400 MHz  PCI-X mmrbc 512 bytes  wire rate throughput of 2.9 Gbit/s  SLAC Dell PCs giving a  Dual 3.0 GHz Xeon CPU FSB 533 MHz  PCI-X mmrbc 4096 bytes  wire rate of 5.4 Gbit/s  CERN OpenLab HP Itanium PCs  Dual 1.0 GHz 64 bit Itanium CPU FSB 400 MHz  PCI-X mmrbc 4096 bytes  wire rate of 5.7 Gbit/s

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 30 Tuning PCI-X: Variation of mmrbc IA32 mmrbc 1024 bytes mmrbc 2048 bytes mmrbc 4096 bytes mmrbc 512 bytes CSR Access PCI-X Sequence Data Transfer Interrupt & CSR Update  byte packets every 200 µs  Intel PRO/10GbE LR Adapter  PCI-X bus occupancy vs mmrbc  Plot:  Measured times  Times based on PCI-X times from the logic analyser  Expected throughput

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester GigEthernet at SC2003 BW Challenge  Three Server systems with 10 GigEthernet NICs  Used the DataTAG altAIMD stack 9000 byte MTU  Send mem-mem iperf TCP streams From SLAC/FNAL booth in Phoenix to:  Pal Alto PAIX  rtt 17 ms, window 30 MB  Shared with Caltech booth  4.37 Gbit hstcp I=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%  Chicago Starlight  rtt 65 ms, window 60 MB  Phoenix CPU 2.2 GHz  3.1 Gbit hstcp I=1.6%  Amsterdam SARA  rtt 175 ms, window 200 MB  Phoenix CPU 2.2 GHz  4.35 Gbit hstcp I=6.9%  Very Stable  Both used Abilene to Chicago

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 32 Summary & Conclusions  Intel PRO/10GbE LR Adapter and driver gave stable throughput and worked well  Need large MTU (9000 or 16114) – 1500 bytes gives ~2 Gbit/s  PCI-X tuning mmrbc = 4096 bytes increase by 55% (3.2 to 5.7 Gbit/s)  PCI-X sequences clear on transmit gaps ~ 950 ns  Transfers: transmission (22 µs) takes longer than receiving (18 µs)  Tx rate 5.85 Gbit/s Rx rate 7.0 Gbit/s (Itanium) (PCI-X max 8.5Gbit/s)  CPU load considerable 60% Xenon 40% Itanium  BW of Memory system important – crosses 3 times!  Sensitive to OS/ Driver updates  More study needed

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 33 PCI Activity: Read Multiple data blocks 0 wait  Read bytes  Each Data block:  Setup CSRs  Data movement  Update CSRs  For 0 wait between reads:  Data blocks ~600µs long take ~6 ms  Then 744µs gap  PCI transfer rate 1188Mbit/s (148.5 Mbytes/s)  Read_sstor rate 778 Mbit/s (97 Mbyte/s)  PCI bus occupancy: 68.44%  Concern about Ethernet Traffic 64 bit 33 MHz PCI needs ~ 82% for 930 Mbit/s Expect ~360 Mbit/s Data transfer CSR Access PCI Burst 4096 bytes Data Block131,072 bytes

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 34 PCI Activity: Read Throughput  Flat then 1/t dependance  ~ 860 Mbit/s for Read blocks >= bytes  CPU load ~20%  Concern about CPU load needed to drive Gigabit link

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 35 BaBar Case Study: RAID Throughput & PCI Activity  3Ware RAID5 parallel EIDE  3Ware forces PCI bus to 33 MHz  BaBar Tyan to MB-NG SuperMicro Network mem-mem 619 Mbit/s  Disk – disk throughput bbcp Mbytes/s (320 – 360 Mbit/s)  PCI bus effectively full! Read from RAID5 Disks Write to RAID5 Disks

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 36 BaBar: Serial ATA Raid Controllers  3Ware 66 MHz PCI  ICP 66 MHz PCI

GNEW2004 CERN March 2004 R. Hughes-Jones Manchester 37 Measure the time between lost packets in the time series of packets sent. Lost 1410 in 0.6s Is it a Poisson process? Assume Poisson is stationary λ(t) = λ Use Prob. Density Function: P(t) = λ e -λt Mean λ = 2360 / s [426 µs] Plot log: slope expect Could be additional process involved VLBI Project: Packet Loss Distribution