Internet2 Network Performance Workshop: Post-Workshop Suggestions

Agenda -- Participants -- Logistical Information -- Presentations -- Post-Workshop Suggestions -- -- Workshop Locations

---

18-19 January 2008
Asia Room
East-West Center
University of Hawaii
Honolulu, HI


Modern science relies on high-performance wide-area networks connecting researchers and resources scattered around the globe.  This workshop focuses on helping scientists, system administrators, and network operators ensure that their local infrastructure supports these scientific demands.  Attendees will now to identify and resolve common problems that cause poor performance over today's networks.

To ensure the best possible outcome from the workshop, Internet2 recommends that the following technical capabilities be met when deploying these tools.

Goal:
This workshop provides scientists, system administrators, and network operators with a standard set of tools that can help identify network performance problems.  Scientists will be able to use these tools to easily and effectively report problems impacting their work.  System administrators will be able to differentiate between host configuration and network infrastructure problems, allowing them to easily resolve the former problems.  Network operators will be able to interpret test results, improving communications with end-users and system administrators, and they will learn how to resolve LAN, MAN, and WAN performance problems.

A specific set of open-source tools (BWCTL, NDT, NPAD, and OWAMP) will be used during this workshop.  Attendees are encouraged to review the web pages for these tools to gain an understanding of these tools. We recommend that participants configure measurement hosts on their own network after the Workshop.

General Requirements:

• Participants should have their own laptop at the workshop to run NDT & NPAD tests. (We will provide a testbed of network and hosts with which to experiment.)
• Participants will receive the most long-term benefit if they incorporate the methods and practices discussed/demoed in this workshop if they install the servers in their own campus network.  The initial step would be to dedicate a single server as the measurement server by booting off the NPToolkit Live-CD.  Once some basic utilization numbers are determined, additional servers, each one dedicated to an individual tool, can be configured and deployed.  Sites should also fill out the PMP form so others can find these new servers.
• Latency measurement can not co-exist with throughput measurements. And different throughput tools can interfere with each other on the same host. We intend to create a meta-scheduler for this purpose but until we do, the easiest solution is to run multiple hosts. This means each measurement point will ideally have 3 hosts currently to run each of the current tools.
• For each measurement point, participants should determine which tools they would like to install. We recommend installing all of the tools at every measurement point even if you need to install them on the same host and only make one of them active at a time for now.
• All of the tools are dependent upon the hosts having good time. Therefore, NTP is a requirement for all hosts. In particular this means:
• Each host should have full access to the internet for NTP (port UDP 123 must be open for sending and receiving). Information on how to configure NTP in a sufficient way for the measurement tools will be one of the topics of the workshop.
• It would be useful for each participant to find several (at least 4 or 5) stratum-1 NTP servers "close" to the measurement points. Ideally these servers will be reached by diverse network paths.

Server Requirements:

The Network Diagnostic Tool (NDT) requires a machine with a Web100-enhanced Linux kernel. We prefer FreeBSD for the machine that runs One-Way Ping (OWAMP), and Linux for the machine that runs the Bandwidth Test Controller. A Web100-enhanced kernel for the Bandwidth Test Controller (BWCTL) machine can be useful, so we recommend that as well. We also recognize that installing and configuring multiple machines is time consuming; if you can only configure one machine before the workshop, make it a web100-enhanced Linux kernel machine.

The Bandwidth Test Controller (BWCTL) has the most stringent hardware requirements, along with NDT. However, the requirements depend on the maximum bandwidth you wish to test. For example, if you have 100 Mbps FastEthernet at your edge, and that is what you want to test, then pretty much any modern PC with a FastEthernet interface will do. Below we will show minimum requirements for Gigabit Ethernet. Machines capable of 10 Gigabit Ethernet are just now becoming available, and will be the most expensive. To see what other sites have deployed, visit the Performance Measurement Point Directory page, and look at any of the "PMP Info" links.

        Hardware:

• CPU: 1+ GHz, 512 KB L2 cache, 133 MHz FSB
• Memory: 512MB minimum, with a motherboard and chipset that supports memory interleaving
• Disk: no stringent requirement, something comfortable. 18 GB is a good minimum value as of early 2005.
• NIC: One-Way Ping only requires FastEthernet. For Gigabit Ethernet, recent motherboard 1000bTX interfaces are sufficient (for example, on Dell rack-mounted servers [e.g. 2650, 1850]). We have used SysKonnect multi-mode fiber interfaces in Abilene; we are told the Intel multi-mode fiber interfaces work best in Dell 2650s.
• Bus: at least 64-bit 66MHz PCI for GigE. (Make sure your NIC and Bus are compatible.)

        OS:

• Linux: Latest kernel version that supports Web100, see the Web100 download page.
• FreeBSD: the current production release should be sufficient, see the FreeBSD site. Note that we haven't tested the latest one just yet; Abilene runs an older version (4.6) and we had some problems starting NTP automatically in release 5.1. That should be fixed in the current release (5.4 as of this writing), but if you have problems let us know (we want to ensure we can use the latest FreeBSD release). You can also try installing the legacy production release (4.11).

        Software:

• C development environment (compiler/libraries/make etc.)
• Web100 libraries/tools
• libpcap library
• we recommend ensuring you have a recent Perl release (5.8.0 or greater), but it is not required to run any of the tools individually.

Note: NDT web100srv process needs root access

One-Way Ping (OWAMP) Requirements:
(These differences are only needed if an actual hardware clock will be connected to the host for better accuracy. This case will not be covered at the workshop unless there are specific requests for it, and extra time is alloted for it. This is a non-trivial exercise.)

        OS:

• FreeBSD 5.3 (NTP PPS kernel turned on in kernel config)        

        NTP:

• UDP 123

        One-Way Ping (OWAMP):

• UDP * (all ephemeral outgoing)
• UDP * (all ephemeral incoming - administrator selectable for upcoming version)
• TCP 4822 (4824 for the upcoming version)

        Bandwidth Test Controller (BWCTL):

• UDP 5001-5010 (iperf ports - admin configurable, but these are recommended for matching with other sites)
• TCP 5001-5010 (iperf ports - admin configurable, but these are recommended for matching with other sites)
• (The "receiver" gets to define the port number, so another reasonable solution is to allow anything outgoing and only a specific set incoming.)
• TCP 4823
• TCP (* outgoing for peer communication between bwctld's)
• TCP (XX incoming - admin configurable, incoming for peer communication between bwctld's. On Abilene we just use ephemeral, we could suggest something I suppose...)

        Network Diagnostic Tool (NDT):

• TCP 7123        (fakewww server default - or real web server)
• TCP 3001-3003 (web100srv ports - testing and analysis engine)