Developing the Web100 Based Network Diagnostic Tool (NDT)

1
Developing the Web100 Based Network Diagnostic
Tool (NDT)

• Performance Analysis
• Tools and Techniques Tutorial
• August 3, 2003
• by
• Richard Carlson

2
Motivation for work

• Develop single shot diagnostic tool that
  doesnt use historical data
• Measure performance to users desktop
• Combine numerous Web100 variables to analyze
  connection
• Develop network signatures for typical network
  problems

3
NDT Benefits

• End-user based view of network
• Doesnt require user to load any new software
• Can identify performance bottlenecks (could be
  host problem)
• Provides some hard evidence to users and
  network administrators to reduce finger pointing

4
Network Signatures

• Duplex Mismatch Detection
• Good results in Campus environment
• Uses multiple Web100 variables to detect
  condition
• Bottleneck Link Type
• Using Packet-pair timing
• Cross-check with Web100 data
• Use TCP-Westwood estimator?

5
Network Signatures

• Faulty Hardware/Link
• Few reports - needs more work
• Link Duplex setting
• Full or half duplex - needs more work
• Normal Congestion
• Is congestion impacting performance? - needs more
  work

6
Network Signatures

• Other common problems?
• Need community input on what problems exist
• Will develop detection signature and update tool
  to report problem

7
Current Deployment

• 3 servers at ANL
• Miranda Externally visible
• Ophelia, Cordelia ANL Internal only
• Non-ANL Servers
• Swiss Education and Research Network (SWITCH)
• University of Michigan - Flint, MI
• University of California - Santa Cruz, CA
• Rochester Institute of Technology - Rochester, NY
• StarLight peering point (coming soon)

8
Availability

• Tools available via anonymous ftp from
  achilles.ctd.anl.gov/pub/web100 directory
• Contains source code and executables
• Email discussion list ltndt_at_anl.govgt
• Majordomo list ltmajordomo_at_achilles.ctd.anl.govgt
• subscribe ndt

9
Installation and Configuration

• Download and Build Web100 kernel/lib
• grab base kernel from ftp.kernel.org
• apply web100 patch
• run favorite kernel config command
• enable experimental code
• enable web100 specific code
• make and install new kernel
• make and install user library web100lib.aso

10
Installation and Configuration

• Download and install the pcap library
• Download from www.tcpdump.org
• Build and install library - libpcap.aso
• needed for link detection algorithm
• Download and install Java SDK
• Download from java.sun.com/products/jdk/1.2/
• use version 1.2.2 for now

11
Installation and Configuration

• Download web100-tools.tar.gz from ANL anonymous
  FTP server (achilles.ctd.anl.gov)
• decide to run pre-compiled or make your own
• ensure web100srv program can access web100lib
  routines
• change LD_LIBRARY_PATH environment variable
• edit /etc/ld.so.conf and add /usr/local/lib, run
  ldconfig
• Reboot new kernel, youre ready to rock roll
• Start fakewww web100srv programs, youre off to
  the races (start.ndt script provided)

12
Results and Observations

• Faulty Hardware identification
• Link Detection algorithm preliminary results
• Mathis et.al formula fails
• Usage statistics
• Demo

13
Effect of Faulty HW Congestion

• Speed
• 94.09 Good
• 22.50 Bad NIC
• 82.66 Bad reverse
• 33.61 Congestion
• 6.99 Good
• 7.15 Bad NIC

• 100 Mbps FD
• Ave Rtt loss loss/sec
• 5.41 0.00 0.03
• 1.38 0.78 15.11
• 6.16 0.00 0.03
• 14.82 0.00 0.10
• 10 Mbps
• 72.80 0.01 0.03
• 8.84 0.75 4.65

14
New Link Detection Algorithm

• Uses Packet-Pair timing
• Small Libpcap program captures data
• Timing taken for each transmit/receive pair
• Results quantized into unique bins
• Statistical analysis on resulting bin counts
• Results compare favorably with Paxsons
  Receiver-Side Estimation Algorithm
• Additional cross-checks begin developed

15
Comparison between old and new link type detection
16
Mathis et.al. Formula fails

• Estimate (K MSS) / (RTT sqrt(loss))
• old-loss (Retrans - FastRetran) / (DataPktsOut
  - AckPktsOut)
• new-loss CongestionSignals / PktsOut
• Estimate lt Measured (K 1)
• old-loss 91/443 (20.54)
• new-loss 35/443 (7.90)
• old agrees with new 26/35 (74.29)

17
Usage by Top Level Domain Name
18
Demo

• http//miranda.ctd.anl.gov7123

19
Disclosure/Disclaimer

• This work was supported (in part) by the Office
  of Science, U.S. Department of Energy under
  Contract W-31-109-ENG-38
• Packet-Pair work was supported by the Cisco
  University Research Program Work-for-Others
  Contract P-03008