NGI White Paper Bill Fink & Pat Gary
March 1997 NASA HPCC/ESS Networking
NASA HPCC/ESS NGI Network Research Agenda
Background
----------
The NASA High Performance Computing and Communications / Earth and
Space Sciences (HPCC/ESS) Networking Group (NH/ENG) is conducting or
planning research in leading edge, high performance network
transmission technologies, prototyping next generation network
protocols to support advanced routing capabilities and the new
Integrated Services Model of the Internet, and testbedding advanced
network applications. These research activities are in support of
the NASA Center for Computational Sciences (NCCS) [1], a major
supercomputer facility serving NASA's ESS researchers, and the
NASA HPCC/ESS [2] Grand Challenge teams which are investigating
fundamental problems in physical and computational sciences.
The HPCC/ESS Computing Testbeds are currently linked via an
OC-12 ATM and HiPPI based Networking Testbed [3] at NASA Goddard to
the OC-12 ATM portion of ATDNet [4], and to an OC-12 ACTS satellite
link [5,6], which then provides OC-12 connectivity to JPL, LeRC, and
the MAGIC [7] testbed. This set of concatenated OC-12 networks is
known as AAMNet (ATDNet-ACTS-MAGIC Network) [8,9,10]. The HPCC/ESS
Network Testbed also has direct OC-3 connectivity to the NASA
NREN [11], the rest of ATDNet, and to Hawaii via ACTS. ATDNet and
ACTS in turn provide OC-3 connectivity to AAI and MAGIC.
Active collaboration is ongoing with our research partners in
the NREN, ATDNet, ACTS, and MAGIC communities across this high
performance network infrastructure, which is enabling the
development of various high performance user applications such as
scientific visualization and collaboration, telemedicine, high
quality video teleconferencing, and the Distr
NGI White Paper NASA HPCC/ESS NGI Network Research Agenda March 1997
NGI Goal 1b: Ultra High Performance Networking
-----------------------------------------------
There are several major challenges that must be met to enable the
attainment and effective use of ultra high performance networks.
This section focuses on the protocol issues which must be solved to
deliver the maximum possible bandwidth to the end user applications.
As bandwidth reaches and exceeds 1 Gbps, the bandwidth*delay
product, which governs the amount of buffering required in the
network and directly affects TCP performance, becomes extremely
large. For example, the OC-12 ACTS satellite path currently has
the largest bandwidth*delay product of 35 MB, and a future cross
country OC-48 ATM network (assuming a RTT of 60 ms) will have a
product of about 15 MB (OC-192 is 62 MB). Also, as the
bandwidth*delay product increases, even small loss rates can
dramatically reduce the effective TCP performance.
* One focus of the NH/ENG research agenda is to better understand
the operation of the standard algorithms in TCP, such as slow
start and congestion avoidance, delayed ACKs, and fast
retransmit and recovery, across very large bandwidth*delay
networks (including satellite nets), to learn how these
algorithms may be tuned for better performance, to study the
effectiveness of the TCP-LFN (RFC1323 TCP window scaling and
time stamp options) and TCP-SACK (RFC2018 TCP selective
acknowledgement option) TCP extensions in optimizing the TCP
throughput, and to determine if any additional TCP extensions
are required to effectively utilize very large bandwidth*delay
networks or if there are any inherent scaling limitations to
the TCP protocol as bandwidth reaches 1 Gbps and beyond.
Interactions with the underlying ATM flow and congestion
control mechanisms, such as explicit rate control in ABR,
will be examined. Other reliable transport protocols, such
as NETBLT and XTP, will be evaluated to determine their
possible applicability, and their performance will be measured
and contrasted with that of TCP-LFN/SACK. Investigate what
changes, if any, will be required to application protocols and
actual user applications.
* As part of this effort, existing tools and methods will be
evaluated, or new tools and methods will be developed, to aid
in performing performance analysis and measurement on these
ultra high performance networks.
Fink/Gary 3/26/97 [Page 2]
NGI White Paper NASA HPCC/ESS NGI Network Research Agenda March 1997
NGI Goal 1a: Scalability of High Performance Networking
--------------------------------------------------------
As the bandwidth requirements are relaxed somewhat, which allows
more participants because of the lower cost, an additional
complexity is introduced, namely all the scaling issues involved
in extending high performance networking to a much larger user
community.
One primary research goal of the NH/ENG is to attempt to create
an effective synergy between the technological capabilities of the
IP(v4/v6) Internet and a global ATM infrastructure, by recognizing
the commonality of functionality and requirements between the IP
and ATM universes, avoiding unnecessary duplication of effort, and
allowing each technology to take maximal advantage of the strengths
of the other.
* The primary scaling issue involves addressing and routing.
Both IPv6 and ATM provide globally unique, hierarchical
network addresses. One approach that would greatly simplify
the linkage between IPv6 and ATM would be to integrate the IPv6
and ATM addressing and routing by simply embedding an IPv6
address in the ATM NSAP address [14] and exporting ATM level
IP routes up to the IP routing infrastructure, an approach known
as the Integrated Routing and Addressing (IRA) Model [15]. It
would have many benefits such as providing direct shortcut
routing at the ATM layer across a hierarchical PNNI ATM network,
eliminate the need for the complexity of NHRP, provide for
distributed ATMARP service, reduce latency for connection setup,
simplify network management, and provide a name service for ATM
NSAP addresses via the DNS. Explore similar possible mappings
for integrating IPv4 and ATM that would also provide a
sufficient level of route aggregation.
* Research other methods of more closely integrating the IP and
ATM layers, such as the IP switching model, which retains the
speed of the ATM switching hardware while discarding the PNNI
routing and UNI signaling protocols, thus effectively
transforming ATM switches into IP routers which support
creation of direct shortcut ATM paths across an underlying
ATM infrastructure.
* Develop methods and tools for simplifying the network management
of IP over ATM networks, to assist with such essential functions
as configuring the network, quickly detecting, isolating, and
fixing problems, and collecting traffic data.
Fink/Gary 3/26/97 [Page 3]
NGI White Paper NASA HPCC/ESS NGI Network Research Agenda March 1997
NGI Goal 2: Implementing the New Integrated Services Model
-----------------------------------------------------------
In addition to providing much higher performance scaled up to a
wide user community, the Next Generation Internet also needs to
support the new Integrated Services Model of the Internet that has
been developed by the IETF specifically to support the requirements
of the new class of real-time applications, including providing
for QOS guarantees and full support of multicast.
* Recommend that the NGI participants form a High-performance
ATM-based MBone (HAMBone), for the purpose of testing real-time
protocols and applications that cannot be tested on the existing
MBone due to bandwidth limitations. This should preferably use
native IP multicasting protocols such as PIM rather than
tunnels. It could initially be IPv4 based, but could later be
expanded to support IPv6. High performance LANs at NGI sites,
such as Fast Ethernet switches, would be connected to the
HAMBone so they could participate in high performance, real-time
multicast sessions. That would allow experience to be gained
with the application of the new Integrated Services Model to
both LAN and WAN environments, for evaluating how effectively
and economically a common infrastructure could support a mix of
both the traditional, elastic applications and the new real-time
applications, and contrast that with the model of providing
separate infrastructures for the different classes of service.
* An integral part of testing the new Integrated Services Model
is experimenting with the resource ReSerVation Protocol (RSVP),
which is the IP layer mechanism for an application to define
its QOS requirements to the network. One area of research
relating to RSVP is the mapping of IP layer RSVP flowspecs to
ATM layer UNI signaling QOS parameters, including evaluating how
well the receiver oriented IP RSVP mechanism can match with the
sender oriented ATM QOS mechanism, and how well the RSVP service
classes (best effort, guaranteed, and predictive) can be mapped
to the ATM traffic classes (CBR, VBR, ABR, and UBR). Another
area of research is the interaction between RSVP and QOS
routing, such as QOSPF.
Fink/Gary 3/26/97 [Page 4]
NGI White Paper NASA HPCC/ESS NGI Network Research Agenda March 1997
References
----------
[1] http://sdcd.gsfc.nasa.gov/NCCS/
[2] http://sdcd.gsfc.nasa.gov/ESS/
[3] http://everest.gsfc.nasa.gov/SCTB/nasanet.gif
[4] http://www.atd.net/ATDNET/
[5] http://kronos.lerc.nasa.gov/acts/acts.htm
[6] http://www.cgrg.ohio-state.edu/other/actsgsn/gsnhome.html
[7] http://www.ukans.magic.net/
[8] http://everest.gsfc.nasa.gov/SCTB/AAMNET_plan.htm
[9] http://www.cgrg.ohio-state.edu/other/actsgsn/aamexp.htm
[10] http://everest.gsfc.nasa.gov/SCTB/aamnet.gif
[11] http://www.nren.nasa.gov/
[12] http://www.nasa.atd.net/hpccess-net-ngi-wp-app.htm
[13] http://www.bell-labs.com/project/MONET/mon_pro.htm
[14] http://www.nasa.atd.net/atm_ipv6ad.htm
[15] http://www.nasa.atd.net/draft-fink-ipatm-ira-00.htm
This NGI White Paper is available at:
http://www.nasa.atd.net/hpccess-net-ngi-wp.htm
A longer form of this NGI White Paper is available at:
http://www.nasa.atd.net/hpccess-net-ngi-wp-long.htm
AUTHORS ADDRESSES
-----------------
Bill Fink Pat Gary
NASA Goddard Space Flight Center NASA Goddard Space Flight Center
Code 933 Code 933
Greenbelt, MD 20771 Greenbelt, MD 20771
Phone: +1 301 286 9423 Phone: +1 301 286 9539
Fax: +1 301 286 1775 Fax: +1 301 286 1634
E-Mail: bill.fink@gsfc.nasa.gov E-Mail: pat.gary@gsfc.nasa.gov
Fink/Gary 3/26/97 [Page 5]