Internet Connectivity and Performance for the HEP Community.

1
Internet Connectivity and Performance for the HEP
Community.

• Presented at HEPNT-HEPiX, October 6, 1999
• by Warren Matthews ltwarrenm_at_slac.stanford.edugt
• Funded by DOE/MICS Internet End-to-end
  Performance Monitoring (IEPM).

2
Goal.

• Why is our project important ?
• Why am I here ?
• What do our studies mean to you ?
• How does it affect HEP research ?

3
Overview.

• Brief review of the monitoring project.
• Performance measured by PingER.
• Trends between Geographical regions.
• A closer look at some events.

4
The Need for Monitoring

• Computing models for BaBar, RHIC, LHC have
  ambitious networking requirements.
• Exabytes or petabytes of data
• analyzed at collaborating Institutes
• Can the Internet perform ?
• Data transfer
• Analysis applications
• Where to allocate resources to make it better.

5
Effect on Applications

• Email
• largely unaffected by poor performance
• File Transfer
• Web (HTML)
• telnet/SSH
• Video/Voice Conferencing
• Even moderate losses can make it unusable

More Interactive and More Sensitive to Problems
6
The PingER Project

• End-to-end
• computers, routers, exchange points
• utilization, routing
• Ping
• ICMP echo request and reply
• 11x100 Byte Pings, 10x1000 Byte Pings

7
PingER Deployment

• 23 Monitoring Sites in 13 Countries.
• 536 Nodes at 381 sites in 55 Countries.
• 2111 pairs.
• 53 Beacon Sites.

8
PingER Metrics

• Packet Loss
• Queue is full
• Round Trip Time
• Path Length
• Speed of Link
• Congestion / Delay at Router

9
PingER and Application Performance

• Same path
• Transit packets are treated the same by routers
• Problems arise if protocols are treated
  differently
• Studies of HTTP show strong correlation
• Better than application performance
• Loss and Response
• Routes
• TTCP lt (MSS / RTTTCP ) (1 / sqrt(p))

10
Interconnected Networks

• Many of the DOE-Funded Labs are connected to the
  Energy Sciences Network (ESnet).
• Many other networks in the U.S.
• vBNS, Abilene I2, THEnet
• European National Research Networks
• GARR, JAnet, Renater, TEN-155

11
Performance on ESnet

• Utilization is currently low
• SLAC upgrade due to anticipated demands of BaBar.
• Packet Loss is negligible.
• Typically lt0.1
• Round Trip Times are good.
• Dominated by path length
• probably the most direct route
• As good as it gets.

12
Probably heavy traffic caused queuing.
13
Bulk transfer - Performance Trends
Bandwidth TCP lt 1460/(RTT sqrt(loss))
14
Performance Between Networks

• Utilization on vBNS and Abilene is low
• Packet Loss is low
• Utilization at Exchange points is also low.
• ESnet lt-gtAbilene, vBNS
• Not the commercial Internet
• No longer one NOC
• Routing Policies

15
CSU change routing to send SLAC packets via
Sacramento.
16
Transatlantic Performance

• Traditional, and well-known, bottleneck.
• Packet Loss can be high.
• Large RTT is unavoidable.

17
(No Transcript)
18
Improving Performance

• Managed Bandwidth
• Quality of Service

19
December
20
Ten-155 became operational on December 11.
To North America
Smurf Filters installed on NORDUnets US
connection.
To Western Europe
21
Regional Performance

• Performance within a NRN is usually good.
• Packet Loss in Italy (GARR) is Good
• Packet Loss in U.K. (JAnet) is Good/Acceptable
• Compare to U.S.
• No exchange point between Laboratories and
  Universities.
• But possibly debate on allocation of resources.

22
General Summary.

• Between Labs,
• Good/Acceptable.
• University Overseas Lab,
• Very Variable (particularly U.S. Universities not
  on ESnet or vBNS).
• Many Connections perform at less than good level,
  particularly remote regions.
• Between Universities,
• Very Variable.

23
Recommendations (Observations)

• Connect to a Research Network.
• Bandwidth with low average utilization.
• Reserved Bandwidth and QoS.

24
Further Information

• CHEP98, TechPub 8178, IEEE
• http//www-iepm.SLAC.Stanford.EDU

Any Questions ?