Some Sample Commercial Network Costs

Download Report

Transcript Some Sample Commercial Network Costs

e-VLBI Development Program
at MIT Haystack Observatory
Alan R. Whitney
Chester A. Ruszczyk
MIT Haystack Observatory
13 July 2005
e-VLBI Workshop
Australia
Current Projects at Haystack Observatory
•
Standardization
–
–
•
Network interfacing equipment for e-VLBI
–
•
Evaluation, development and deployment of monitoring systems
Intelligent Applications
–
–
–
•
Mark 5 VLBI data system
Network Monitoring
–
•
VSI-E Draft VSI-E standard distributed in January 2004
Reference implementation released in October 2004
Automation of e-VLBI transfers an ongoing process
Development of optimization-based algorithms for intelligent
applications ongoing (EGAE)
Intelligent optically-switched networks (DRAGON)
e-VLBI Experiments
–
Goal to put e-VLBI into routine use
VSI-E Architecture
VSI-E
•
Purpose:
–
•
Characteristics:
–
–
–
–
–
•
To specify standardized e-VLBI data formats and transmission protocols that
allow data exchange between heterogeneous VLBI data systems
Based on standard RTP/RTCP high-level protocols
Allows choice of IP transport protocols (TCP-IP, UDP, FAST, etc.)
Scalable Implementation; supports up to 100Gbps
Ability to transport individual data-channel streams as individual packet
streams; potentially useful for distributed correlators
Ability to make use of multicasting to transport data and/or control information
in an efficient manner
Status
–
–
–
–
Draft VSI-E specification completed January 2004
Prototype VSI-E prototype implementation Nov 2004
Practical implementation for K5 and Mark 5 now is progress
Plan to use VSI-E in real-time demo at SC05, Nov 05
Reaching 1024 Mbps with Mark 5
• Achieving 1024Mbps with Mark 5 is challenging
• Can move ~1.2 Gbps between StreamStor card memory via PCI bus,
but
– If GigE NIC is on same PCI bus, bus contention slows aggregate transfers
to ~400-550Mbps, depending on motherboard
– Single GigE connections tops out at ~980Mbps (theoretically and
experimentally)
– Typical GigE drivers require interrupt service every Ethernet frame; can
generate up to ~100,000 interrupts/sec
• Elements of Solution
–
–
–
–
Capable motherboard with multiple independent PCI buses
Dual ‘channel-bonded’ GigE links
Driver or hardware interrupt mitigation; use of ‘jumbo frames’
Careful software structure
Mark 5 e-VLBI Connectivity
• Mark 5 supports a triangle of connectivity for e-VLBI requirements
Disc array
Data Port/FPDP
PCI bus/Network
(64bit/66MHz)
Mark 5 can support several possible e-VLBI modes:
• e-VLBI data buffer (first to Disc Array, then to Network); vice versa
• Direct e-VLBI (Data Port directly to Network); vice versa
• Data Port simultaneously to Disc Array and Network at ~800 Mbps
Anatomy of a (fairly) modern motherboard
(Tyan Thunder i7501 Pro)
2.8GHz
Xeon
2.8GHz
Xeon
FSB 64-bit 533MHz (4.2GB/sec)
FPDP
Bus
Stream
Stor
64-bit 66MHz
Dual
GigE
PCI-X
64-bit
133MHz
PCI-X
Intel 82546EB 64-bit
Jumbo frames 133MHz
to 16kB;
Interrupt mitigation;
channel-bonding
PCI
Bridge
HubLink 2.0
1.6GB/s
Intel P64H2
PCI
Bridge
Intel P64H2
Memory
Control
Hub
1GB
memory
64-bit
133MHz
Dual-edge
1GB
memory
HubLink 2.0
1.6GB/s
HubLink 1.0
I/O
Cntlr
PCI
32-bit
33MHz
Mark 5A
I/O
Board
PC2100
266MHz
Best transfer rates to date
• Memory-to-memory transfers between Tyan motherboards – ~1900Mbps
• Uses dual channel-bonded GigE connection
• Mark 5A-to-memory transfer – ~1200 Mbps
• Required major re-working of Mark 5A software to improve efficiency of
data-transfer to/from NIC, minimize number of internal buffer-to-buffer
transfers, and support multiple threads
• More work still to be done to achieve routine 1024 Mbps Mark5-toMark5 transfers
• We plan to concentrate our efforts on implementing and optimizing with
VSI-E to achieve 1024 Mbps
• There should be no performance difference between Mark 5A
and Mark 5B
e-VLBI Network Monitoring
• Use of centralized/integrated network monitoring helped to enable
identification of bottleneck (hardware fault)
• Automated monitoring allows view of network throughput variation
over time
– Highlights route changes, network outages
• Automated monitoring also helps to highlight any throughput issues at
end points:
– E.g. Network Inteface Card failures, Untuned TCP Stacks
• Integrated monitoring provides overall view of network behavior at a
glance
• Also examining performance-monitoring packages such as MonaLisa,
which would provide better standardization
Network State DataBase (NSDB)
•
Tool to keep track of state of e-VLBI state:
–
–
–
•
Network performance
Configuration of end systems
State of end systems
Integrates and builds on standard monitoring tools to provide a
single, coherent view of e-VLBI network state:
–
–
–
Maintain continuous state monitoring of entire e-VLBI system
Essential for being able to identify issues with network/end system
configuration
Diagnose at-a-glance (cf. current practice)
NSDB Architecture
e-VLBI Weather Map Web Page
(Haystack to Kashima)
http://web.haystack.mit.edu/e-vlbi/evlbi.html
Network Layer Statistics
New Application-Layer Protocols for e-VLBI
• Based on observed usage statistics of networks such as Abilene, it is
clear there is much unused capacity
• New protocols are being developed which are tailored to e-VLBI
characteristics; for example:
– Can tolerate some loss of data (perhaps 1% or so) in many cases
– Can tolerate delay in transmission of data in many cases
• ‘Experiment-Guided Adaptive Endpoint’ (EGAE) strategy being
developed at Haystack Observatory under 3-year NSF grant:
– Will ‘scavenge’ and use ‘secondary’ bandwidth
– ‘Less than best effort’ service will not interfere with high-priority users
– Translates science-user criteria into network constraints
Automation of e-VLBI transfers
•
•
Based on EGAE, major effort is now underway to fully automate
routine e-VLBI file transfers
Algorithms are being built around use of standardized e-VLBI filenaming conventions (as agreed by Himwich, Koyama, Reynolds,
Whitney, Nov 2004); see memo #49 at ftp://web.haystack.edu/pub/evlbi/memoindex.html
–
We urge universal adoption of standardized e-VLBI file naming for
ease of data interchange
Experimental and Production e-VLBI
•
August 2004:
–
–
Haystack link link upgraded to 2.5 Gbps
Real-time fringes at 128 Mbps, Westford and GGAO antennas, Haystack
Correlator
•
September 2004:
–
•
November 2004
–
–
•
Real-time fringes Westford-Onsala at 256Mbps
Used optically-switched light paths over part of route
October 2004 – present
–
•
Real-time e-VLBI demonstration at SC2004 at 512 Mbps
Use DRAGON optically-switched light paths
February 2005
–
–
•
Real-time fringes at 512 Mbps, Westford and GGAO antennas, Haystack
Correlator
Regular transfers from Kashima (~300GB per experiment; ~200 Mbps)
Starting April 2005
–
–
Routine weekly transfers from Tsukuba (~1.2TB/transfer)
Preparing for CONT05 (15 days continuously; ~1TB/day)
Real-time e-VLBI SC2004 Demo
Haystack
Westford
Bossnet
Pittsburgh Convention Center
DRAGON
Goddard
GGAO
DRAGON Project
(Dynamic Resource Allocation for FMPLS Optical Networks)
•
Dynamically-provisionally optically-switched network research
project
–
•
•
10GBPS DRAGON network is being installed around Washington,
D.C. area, with connections to Abilene, HOPI and NLR
e-VLBI is primary demonstration application, using 2.4Gbps
dedicated connection to Haystack
–
–
•
U. of Maryland, ISI – PI’s
Programmatic interfaces to EGAE are under development
Hope to upgrade Haystack connection to 10 Gbps in near future
DRAGON will play a prominent role in e-VLBI demos scheduled for
iGRID (Sep 05) and SC05 (Nov 05)
Abilene
ISIE
M10
WXC
EXC2
RE2
RE1
l
HAYS
ATDnet/
Bossnet
DRAGON Network
ARLG
EXC1
MCLN
WXC2
WXC2
RE4
RE1
l
l
OSPF control plane adjacencies
l
l
WXC1
RE3
RE3
GSFC
NCSA
l
RE1
RE1
UMCP
CLPK
HOPI
Movaz Networks
iWSS Optical Switch
•
•
•
•
MEMS-based switching fabric
400 x 400 wavelength switching, scalable to 1000s x 1000s
9.23"x7.47"x3.28" in size
Integrated multiplexing and demultiplexing, eliminating the cost
and challenge of complex fiber management

Dynamic power equalization (<1 dB uniformity),
eliminating the need for expensive external equalizers
Ingress and egress fiber channel monitoring outputs to
provide sub-microsecond monitoring of channel
performance using the OPM
Switch times < 5ms


In summary - Some lessons learned
• High-performance e-VLBI is still hard to do
–
–
–
–
–
Cannot count on consistent performance
Varying traffic loads
Network configuration changes
Equipment failures
Continuous network monitoring is critical to success of on-demand RT eVLBI
• Jumbo-frame support is important at rates >~256Mbps on GigE
– Jumbo-frame support is spotty, but improving
Some Challenges
•
•
Network bottlenecks well below advertised rates
Performance of transport protocols
–
•
Throughput limitations of COTS hardware
–
•
•
Disk-I/O - Network
Complexity of e-VLBI experiments
–
•
•
untuned TCP stacks, fundamental limits of regular TCP
e-VLBI experiments currently require significant network expertise to
conduct
Time-varying nature of network
Define standard formats for transfer of data and control information
between different VLBI systems
‘Last-mile’ connectivity to telescopes
–
–
Most telescopes are deliberately placed in remote areas
Extensive initiatives in Europe, Japan and Australia to connect;
U.S. is lagging
Some Frustrations
• Telescope connectivity, particularly in U.S. , remains a significant
challenge
–
–
–
–
–
–
–
–
Westford – 1 Gbps
GGAO – 1 Gbps
Arecibo – 155 Gbps
VLBA – not connected
GBT – not connected
CARMA – not connected
JCMT – not connected
SMA – not connected
• Much difficulty in securing funding support from NSF Astronomy for eVLBI
– Need to develop convincing science case
Future Directions
•
•
•
•
•
•
Further EGAE and VSI-E development and deployment
Improved IP protocols for e-VLBI
Optically-switched networks for highly provisioned high-data-rate
pipes
Solving ‘last mile’ problem to U.S. telescopes
Distributed correlation using clusters and/or highly distributed PC’s
Extending to higher bandwidths
–
–
•
Haystack has Astronomy NSF grant to push for 4Gbps/station
Preparing NSF proposal to extend to 16Gbps/station using new digitalfilter and recording technology
Continuing to move e-VLBI into routine practice on a global basis
e-VLBI Technical Working Group
• Established at this e-VLBI workshop as group of technical experts,
David Lapsley chair
• On hold until David Lapsley replacement is on-board
• Hope to re-invigorate at July e-VLBI workshop in Sydney
• Objectives
– Evaluate e-VLBI/VSI-E hardware/software/procedures
– Implement standardized global e-VLBI network
performance/monitoring tools
– Provide expert assistance to e-VLBI users
• ~2 members from each major e-VLBI geographical area
Thank you - THE END
Questions?
Antenna/Correlator Connectivity
•
•
•
•
•
•
•
•
•
•
•
•
•
•
JIVE Correlator (6 x 1 Gbps)
Haystack (2.5 Gbps)
Kashima, Japan (1 Gbps)
Tsukuba, Japan (1 Gbps)
GGAO, MD (10 Gbps)
Onsala, Sweden (1 Gbps)
Torun, Poland (1 Gbps)
Westerbork, The Netherlands (1 Gbps)
Westford, MA (2 Gbps)
Jodrell Bank (1 Gbps?)
Arecibo, PR (155 Mbps)
Wettzell, Germany (~30 Mbps)
Kokee Park, HA (nominally ~30 Mbps, but problems)
TIGO (~2 Mbps)
In progress:
• Australia – plan to connect all major antennas at 10Gbps!
• Hobart – agreement reached to install high-speed fiber
• NyAlesund – work in progress to provide ~200Mbps link to
NASA/GSFC