The Intersection of Grids and Networks: Where the Rubber Hits the Road
William E. Johnston ESnet Manager and Senior Scientist Lawrence Berkeley National Laboratory
1
Objectives of this Talk
• How a production R&E network works • Why some types of services needed by Grids / widely distributed computing environments are hard
2
Outline
• • •
•
•
How do Networks Work? Role of the R&E Core Network ESnet as a Core Network o
ESnet Has Experienced Exponential Growth Since 1992
o
ESnet is Monitored in Many Ways
o
How Are Problems Detected and Resolved?
Operating Science Mission Critical Infrastructure o
Disaster Recovery and Stability
o
Recovery from Physical Attack / Failure
o
Maintaining Science Mission Critical Infrastructure in the Face of Cyberattack
Services that Grids need from the Network o
Public Key Infrastructure example 3
How Do Networks Work?
• Accessing a service, Grid or otherwise, such as a Web server, FTP server, etc., from a client computer and client application (e.g. a Web browser_ involves o
Target host names
o
Host addresses
o
Service identification
o
Routing
4
How Do Networks Work?
• When one types “google.com” into a Web browser to use the search engine, the following takes place o
The name “google.com” is resolved to an Internet address by the Domain Name System (DNS) – a hierarchical directory service
o
The address is attached to a network packet (which carries the data – a google search request in this case) which is then sent out of the computer into the network
o
The first place that the packet reaches is a router that must decide how to get that packet to its desitnatiion (google.com)
5
How Do Networks Work? o
In the Internet, routing is done “hot potato” - Routers are in your site LANs and at your ISP, and each router typically communicates directly with several other routers - The first router to receive your packet takes a quick look at the address and says, if I send this packet to router B that will probably take it closer to its destination. So it sends it to B without further adieu. - Router B does the same thing, and so forth, until the packet reaches google.com
o
What makes this work is routing protocols that exchange reachability information between all directly connected routers – “BGP” is the most common such protocol in WANs 6
How Do Networks Work?
• Once the packet reaches its destination (the computer called google.com) it must be delivered to the google search engine, as opposed to the google mail server that may be running on the same machine. o
This is accomplished with a service identifier that is put on the packet by the browser (the client side application) - The service identifier says that this packet is to be delivered to the Web server on the destination system – on each system every server/service has a unique identified called a “port number”
o
So when someone says that the Blaster/Lovsan worm is attacking port 135 on the system called google.com, they mean that a worm program somewhere in the Internet is trying to gain access to the service at port 135 on google.com (usually to exploit a vulnerability). 7
Role of the R&E Core Network: Transit (Deliver Every Packet) LBNL
router
core router
router
ESnet (Core network)
border router
gateway router
core routers •focus on highspeed packet forwarding
core router
border/gateway routers •implement separate site and network provider policy (including site firewall policy)
peering router
peering routers •implement/enforce routing policy for each provider •provide cyberdefense
peering router
Big ISP (e.g. SprintLink)
router
router
router router router
router
Google, Inc. 8
Outline
• • •
•
•
How do Networks Work? Role of the R&E Core Network ESnet as a Core Network o
ESnet Has Experienced Exponential Growth Since 1992
o
ESnet is Monitored in Many Ways
o
How Are Problems Detected and Resolved?
Operating Science Mission Critical Infrastructure o
Disaster Recovery and Stability
o
Recovery from Physical Attack / Failure
o
Maintaining Science Mission Critical Infrastructure in the Face of Cyberattack
Services that Grids need from the Network o
Public Key Infrastructure example 9
What is ESnet
• ESnet is a large-scale, very high bandwidth network providing connectivity between DOE Science Labs and their science partners in the US, Europe, and Japan
• Essentially all of the national data traffic supporting US open science is carried by two networks – ESnet and Internet-2 / Abilene (which plays a similar role for the university community)
• ESnet is very different from commercial ISPs (Internet Service Providers) like Earthlink, AOL, etc. o o
Most big ISPs provide small amounts of bandwidth to a large number of sites ESnet supplies very high bandwidth to a small number of sites 10
ESnet Connects DOE Facilities and Collaborators CA*net4 KDDI (Japan) France Switzerland Taiwan (TANet2)
Australia CA*net4 Taiwan (TANet2) Singaren
GEANT - Germany - France - Italy - UK - etc. Sinet (Japan) Japan – Russia(BINP)
W PN
CA*net4 CERN MREN Netherlands Russia StarTap Taiwan (ASCC)
G
SEA HUB LIGO
ne
ESnet IP
ne
LBNL NERSC SLAC
QWEST ATM
LLNL
AP
ne FNAL
ANL-DC INEEL-DC ORAU-DC
ANL
LLNL/LANL-DC
AMES
SNV HUB
MAE-W Fix-W
Ab ile
Ch iN
L
CHI HUB
YUCCA MT
MAE-E
4xLAB-DC
KCP
LANL
SDSC
GA 42 end user sites Office Of Science Sponsored (22)
SNLA Allied Signal DOE-ALB
PA NT E
NOAA SRS
Abilene X
HUB ELP
NNSA Sponsored (12) Joint Sponsored (3) Other Sponsored (NSF LIGO, NOAA) Laboratory Sponsored (6) peering points SNV HUB ESnet hubs
JLAB
ORAU
OSTI ARM
ALB HUB
PAIX-E
DC HUB
ORNL
MIT
BNL NY-NAP PPPL
GTN&NNSA
L TE CH E B
L NRE
W IXPA ix qin Eu
INEE
SNLL
B U H
JGI
ile Ab
C
t
TW
YC N
ligh Star
Japan
Ab ile
PNNL
ESnet core ring: Packet over SONET Optical Ring and Hubs
ATL HUB
International (high speed) OC192 (10G/s optical) OC48 (2.5 Gb/s optical) Gigabit Ethernet (1 Gb/s) OC12 ATM (622 Mb/s) OC12 OC3 (155 Mb/s) T3 (45 Mb/s) T1-T3 T1 (1 Mb/s)
11
Current Architecture ESnet site
site LAN
Site IP router ESnet hub
RTR
ESnet IP router
RTR
10GE
• usually SONET data framing or Ethernet data framing • can be clear digital channels (no framing – e.g. for digital HDTV)
RTR
10GE
Lambda channels are converted to electrical channels
ESnet core
Site – ESnet network policy demarcation (“DMZ”)
Wave division multiplexing • today typically 64 x 10 Gb/s optical channels per fiber • channels (referred to as “lambdas”) are usually used in bi-directional pairs
A ring topology network is inherently reliable – all single point failures are mitigated by routing traffic in the other direction around the ring. RTR
RTR
optical fiber ring RTR 12
Peering – ESnet’s Logical Infrastructure – Connects the DOE Community With its Collaborators Australia CA*net4 Taiwan (TANet2) Singaren
PNW-GPOP
CA*net4 CERN MREN Netherlands Russia StarTap Taiwan (ASCC)
KDDI (Japan) France
GEANT - Germany - France - Italy - UK - etc SInet (Japan) KEK Japan – Russia (BINP)
SEA HUB
2 PEERS
CH
LBNL
CalREN2
39 PEERS 3 PEERS
FIX-W
PAIX-W
CENIC SDSC
B HU V Abilene 2 PEERS SN MAE-W
IN
AP
Abilene + 7 Universities
ST
Japan
AR L
IG H
T
Distributed 6TAP 19 Peers
Abilene 1 PEER
NYC HUBS
1 PEER
NYNAP
5 PEERS 26 PEERS
MAX GPOP 22 PEERS
20 PEERS EQX-SJ GA
University International Commercial
EQ XA
SH
LANL TECHnet
Commercial
ESnet Peering (connections to other networks)
6 PEERS
MAE-E PAIX-E
Abilene
ATL HUB
ESnet provides complete access to the Internet by managing the full complement of Global Internet routes (about 150,000) at 10 general/commercial peering points + high-speed peerings w/ Abilene and the international networks.
What is Peering?
•
• •
Peering points exchange routing information that says “which packets I can get closer to their destination” ESnet daily peering report (top 20 of about 100) This is a lot of work
peering with this outfit is not random, it carries routes that ESnet needs (e.g. to the Russian Backbone Net)
AS
routes
peer
1239
63384
SPRINTLINK
701
51685
UUNETALTERNET
209
47063
QWEST
3356
41440
LEVEL3
3561
35980
CABLEWIRELESS
7018
28728
ATT-WORLDNET
2914
19723
VERIO
3549
17369
GLOBALCENTER
5511
8190
OPENTRANSIT
174
5492
COGENTCO
6461
5032
ABOVENET
7473
4429
SINGTEL
3491
3529
CAIS
11537
3327
ABILENE
5400
3321
BT
4323
2774
TWTELECOM
4200
2475
ALERON
6395
2408
BROADWING
2828
2383
XO
7132
1961
SBC
14
What is Peering?
•
Why so many routes? So that when I want to get to someplace out of the ordinary, I can get there. For example: http://www-sbras.nsc.ru/eng/sbras/copan/microel_main.html (Technological Design Institute of Applied Microelectronics of SB RAS 630090, Novosibirsk, Russia) Peering routers Start: 134.55.209.5
snv-lbl-oc48.es.net
ESnet core
134.55.209.90
snvrt1-ge0-snvcr1.es.net
ESnet peering at Sunnyvale
63.218.6.65
pos3-0.cr01.sjo01.pccwbtn.net
AS3491 CAIS Internet
63.218.6.38
pos5-1.cr01.chc01.pccwbtn.net
“
“
63.216.0.53
pos6-1.cr01.vna01.pccwbtn.net
“
“
63.216.0.30
pos5-3.cr02.nyc02.pccwbtn.net
“
“
63.218.12.37
pos6-0.cr01.ldn01.pccwbtn.net
“
“
63.218.13.134
rbnet.pos4-1.cr01.ldn01.pccwbtn.net
AS3491->AS5568 (Russian Backbone Network) peering point
195.209.14.29
MSK-M9-RBNet-5.RBNet.ru
Russian Backbone Network
195.209.14.153
MSK-M9-RBNet-1.RBNet.ru
“
“
195.209.14.206
NSK-RBNet-2.RBNet.ru
“
“
Finish: 194.226.160.10
Novosibirsk-NSC-RBNet.nsc.ru
RBN to AS 5387 (NSCNET-2)
15
ESnet is Engineered to Move a Lot of Data
ESnet is currently transporting about 250 terabytes/mo.
ESnet Monthly Accepted Traffic
300
TBytes/Month
250 200 150 100
Annual growth in the past five years has increased from 1.7x annually to just over 2.0x annually.
50
2004
2003
2002
2001
2000
1999
1998
1997
1996
1995
1994
1993
1992
1991
1990
0
16
Who Generates Traffic, and Where Does it Go? ESnet Inter-Sector Traffic Summary, Jan 2003 / Feb 2004 (1.7X overall traffic increase, 1.9X OSC increase) (the international traffic is increasing due to BABAR at SLAC and the LHC tier 1 centers at FNAL and BNL)
72/68%
DOE sites
DOE is a net supplier of data because DOE facilities are used by universities and commercial entities, as well as by DOE researchers
21/14% ESnet
~25/18%
14/12% 17/10% 10/13%
Note that more that 90% of the ESnet traffic is OSC traffic ESnet Appropriate Use Policy (AUP) All ESnet traffic must originate and/or terminate on an ESnet an site (no transit traffic is allowed)
R&E (mostly universities) Peering Points
53/49% DOE collaborator traffic, inc. data
Commercial
9/26%
International
4/6% Traffic coming into ESnet = Green Traffic leaving ESnet = Blue Traffic between sites % = of total ingress or egress traffic
17
1 terabyte/day
ESnet Top 20 Data Flows, 24 hrs., 2004-04-20
S
C A L
) S (U
→
2 IN
P3
R) F (
→ ) S
N R CE
va d a
) I( T
A small number of science users account for a significant fraction of all ESnet traffic
) S U US) ( N o b( F (U g ) b IN ica mila R) S h ila U C ) → er 3 (F C ( m . r S F ) U e S F → N2P SLA C (U ) ) → ) (U I K ) ) ) A A S (C K (U US ) C → SL IT → ( U ) U b A ) ) ( a ab T ( ET l3 ( NL S va IT to US b L ( → SL n a ) a E E L NE AN ve et ( (U Pad o F( il r v E rm T o A DO DO JA → J Le Fn LAC FN ad (D e . B P F → UR S IN → → U CE iN → S) N ) ) b W F a ab U US S ) → S) → S ( L N IN (U ab e ( e → FR (U L DF OE E ( C il D DO LA rm onn onn P3 ilab g g 2 m S Fe Ar Ar IN Fer
P
18
Top 50 Traffic Flows Monitoring – 24hr – 1 Int’l Peering Point
10 flows > 100 GBy/day More than 50 flows > 10 GBy/day
19
Scalable Operation is Essential
• R&E networks typically operate with a small staff • The key to everything that the network provides is scalability o
How do you manage a huge infrastructure with a small number of people?
o
This issue dominates all others when looking at whether to support new services (e.g. Grid middleware) - Can the service be structured so that its operational aspects do not scale as a function of the use population? - If not, then it cannot be offered as a service
20
Scalable Operation is Essential The entire ESnet network is operated by fewer than 15 people
Infrastructure (6 FTE)
Core Engineering Group (5 FTE) 7X24 On-Call Engineers (7 FTE) 7X24 Operations Desk (2-4 FTE)
Science Services (middleware and collaboration tools) (5 FTE)
Management, resource management, circuit accounting, group leads (4 FTE)
•
21
•Automated, real-time monitoring of traffic levels and operating state of some 4400 network entities is the primary network operational and diagnosis tool Performance
Hardware Configuration
Network Configuration
SecureNet
OSPF Metrics (internal routing and connectivity)
IBGP Mesh (WAN routing and connectivity)
How Are Problems Detected and Resolved? Australia CA*net4 Taiwan (TANet2) Singaren
CA*net4 When a hardware KDDI (Japan) alarm goes off France Switzerland here, the 24x7Taiwan (TANet2) operator is notified
CA*net4 CERN MREN Netherlands Russia StarTap Taiwan (ASCC)
GEANT - Germany - France - Italy - UK - etc Sinet (Japan) Japan – Russia(BINP) Nevis Yale
SEA HUB
LIGO
de an Br
PNNL
YC N
is
ESnet IP
Japan
B U H
MIT
TW
JGI
LBNL NERSC SLAC
C
INEE
SNLL
L
QWEST ATM
LLNL
ANL
LLNL/LANL-DC
AMES
SNV HUB
CHI HUB
BNL PPPL
4xLAB-DC GTN&NNSA
L TE
Allied Signal
L NRE
CH BE YUCCA MT
OSTI ARM
ALB HUB
DC HUB
ORNL
LANL
SDSC
GA
FNAL
ANL-DC INEEL-DC ORAU-DC
SNLA Allied Signal DOE-ALB
PA NT E
HUB ELP
JLAB
ORAU
NOAA SRS
X
ATL HUB
International (high speed) OC192 (10G/s optical) OC48 (2.5 Gb/s optical) Gigabit Ethernet (1 Gb/s) OC12 ATM (622 Mb/s) OC12 OC3 (155 Mb/s) T3 (45 Mb/s) T1-T3 T1 (1 Mb/s)
23
ESnet is Monitored in Many Ways Performance
Hardware Configuration
ESnet configuration
SecureNet
OSPF Metrics
IBGP Mesh
24
Drill Down into the Configuration DB to Operating Characteristics of Every Device
e.g. cooling air temperature for the router chassis air inlet, hot-point, and air exhaust for the ESnet gateway router at PNNL
Problem Resolution
• Let’s say that the diagnoistics have pinpointed a bad module in a router rack in the ESnet hub in NYC
• Almost all high-end routers, and other equipment that ESnet uses, have multiple, redundant modules for all critical functions
• Failure of a module (e.g. a power supply or a control computer) can be corrected on-the-fly, without turning off the power or impacting the continued operation of the router
• Failed modules are typically replaced by a “smart hands” service at the hubs or sites o
One of the many essential scalability mechanisms 26
ESnet is Monitored in Many Ways Performance
Hardware Configuration
ESnet configuration
SecureNet
OSPF Metrics
IBGP Mesh
27
Drill Down into the Hardware Configuration DB for Every Wire Connection
Equipment rack detail at AOA, NYC Hub (one of the 10 Gb/s core optical ring sites)
The Hub Configuration Database
• Equipment wiring detail for two modules at the AOA, NYC Hub • This allows “smart hands” – e.g., Qwest personnel at the NYC site – to replace modules for ESnet)
What Does this Equipment Actually Look Like?
Picture detail
Equipment rack detail at NYC Hub, 32 Avenue of the Americas (one of the 10 Gb/s core optical ring sites) 30
Typical Equipment of an ESnet Core Network Hub Sentry power 48v 30/60 amp panel ($3900 list) Sentry power 48v 10/25 amp panel ($3350 list) DC / AC Converter ($2200 list)
Lightwave Secure Terminal Server ($4800 list)
Juniper M20 AOA-PR1 (peering RTR) ($353,000 list)
Qwest DS3 DCX
AOA Performance Tester ($4800 list)
Cisco 7206 AOA-AR1 (low speed links to MIT & PPPL) ($38,150 list)
ESnet core equipment @ Qwest 32 AofA HUB NYC, NY (~$1.8M, list)
Juniper OC192 Optical Ring Interface (the AOA end of the OC192 to CHI ($195,000 list)
Juniper T320 AOA-CR1 (Core router) ($1,133,000 list)
Juniper OC48 Optical Ring Interface (the AOA end of the OC48 to DC-HUB ($65,000 list) 31
Outline
• • •
•
•
How do Networks Work? Role of the R&E Core Network ESnet as a Core Network o
ESnet Has Experienced Exponential Growth Since 1992
o
ESnet is Monitored in Many Ways
o
How Are Problems Detected and Resolved?
Operating Science Mission Critical Infrastructure o
Disaster Recovery and Stability
o
Recovery from Physical Attack / Failure
o
Maintaining Science Mission Critical Infrastructure in the Face of Cyberattack
Services that Grids need from the Network o
Public Key Infrastructure example 32
Operating Science Mission Critical Infrastructure
•
ESnet is a visible and critical piece of DOE science infrastructure o
•
if ESnet fails,10s of thousands of DOE and University users know it within minutes if not seconds
Requires high reliability and high operational security in the systems that are integral to the operation and management of the network o
Secure and redundant mail and Web systems are central to the operation and security of ESnet - trouble tickets are by email - engineering communication by email - engineering database interfaces are via Web
o
Secure network access to Hub routers
o
Backup secure telephone modem access to Hub equipment
o
24x7 help desk and 24x7 on-call network engineer
[email protected] (end-to-end problem resolution)
33
Disaster Recovery and Stability SEA HUB
LBNL SNV HUB
TWC
Remote Engineer • partial duplicate infrastructure
Engineers, 24x7 Network Operations Center, generator backed power • Spectrum (net mgmt system) • DNS (name – IP address translation) • Eng database • Load database • Config database • Public and private Web • E-mail (server and archive) • PKI cert. repository and revocation lists • collaboratory authorization ALB HUB service
Remote Engineer • partial duplicate infrastructure
DNS AMES
BNL
CHI HUB
NYC HUBS
PPPL DC HUB
Remote Engineer
AT
UB EL P H
LH
UB
Duplicate Infrastructure Currently deploying full replication of the NOC databases and servers and Science Services databases in the NYC Qwest carrier hub
• The network must be kept available even if, e.g., the West Coast is disabled by a massive earthquake, etc. Reliable operation of the network involves • remote Network Operation Centers (3) • replicated support infrastructure • generator backed UPS power at all critical network and infrastructure locations
• high physical security for all equipment • non-interruptible core - ESnet core operated without interruption through o o o
N. Calif. Power blackout of 2000 the 9/11/2001 attacks, and the Sept., 2003 NE States power blackout
34
Recovery from Physical Attack / Core Ring Failure normal traffic flow Chicago (CHI)
The Hubs have lots of connections (42 in all)
Sunnyvale (SNV)
New York (AOA)
X
break in the ring
reversed traffic flow Washington, DC (DC)
ESnet backbone (optical fiber ring)
Hubs (backbone routers and local loop connection points)
El Paso (ELP)
Atlanta (ATL)
We can route traffic either way around the ring, so any single failure in the ring is transparent to ESnet users
The local loops are still single points of failure
Local loop (Hub to local site) ESnet border router
DMZ
Site gateway router
Site LAN
Site 35
Maintaining Science Mission Critical Infrastructure in the Face of Cyberattack
•
A Phased Security Architecture is being implemented to protects the network and the ESnet sites
•
The phased response ranges from blocking certain site traffic to a complete isolation of the network which allows the sites to continue communicating among themselves in the face of the most virulent attacks o
Separates ESnet core routing functionality from external Internet connections by means of a “peering” router that can have a policy different from the core routers
o
Provide a rate limited path to the external Internet that will insure siteto-site communication during an external denial of service attack
o
Provide “lifeline” connectivity for downloading of patches, exchange of e-mail and viewing web pages (i.e.; e-mail, dns, http, https, ssh, etc.) with the external Internet prior to full isolation of the network
Cyberattack Defense ESnet first response – filters to assist a site
ESnet second response – filter traffic from outside of ESnet
ESnet third response – shut down the main peering paths and provide only limited bandwidth paths for specific “lifeline” services
X
X router
ESnet
peering router
router
LBNL X ¾Lab first response – filter incoming traffic at their ESnet gateway router
gateway router
border router
attack traffic
router
peering router
border router
Lab
¾Sapphire/Slammer worm infection created a Gb/s of traffic on the ESnet core until filters were put in place (both into and out of sites) to damp it out.
Lab
gateway router
37
ESnet WAN Security and Cybersecurity
•
Cybersecurity is a new dimension of ESnet security Security is now inherently a global problem o As the entity with a global view of the network, ESnet has an important role in overall security o
30 minutes after the Sapphire/Slammer worm was released, 75,000 hosts running Microsoft's SQL Server (port 1434) were infected. (“The Spread of the Sapphire/Slammer Worm,” David Moore (CAIDA & UCSD CSE), Vern Paxson (ICIR & LBNL), Stefan Savage (UCSD CSE), Colleen Shannon (CAIDA), Stuart Staniford (Silicon Defense), Nicholas Weaver (Silicon Defense & UC Berkeley EECS) http://www.cs.berkeley.edu/~nweaver/sapphire ) Jan., 2003 38
ESnet and Cybersecurity Sapphire/Slammer worm infection hits creating almost a full Gb/s (1000 megabit/sec.) traffic spike on the ESnet backbone
39
Outline
• • • •
Role of the R&E Transit Network ESnet is Driven by the Requirements of DOE Science Terminology – How Do Networks Work? How Does it Work? – ESnet as a Backbone Network ESnet Has Experienced Exponential Growth Since 1992 o ESnet is Monitored in Many Ways o How Are Problems Detected and Resolved? o
•
Operating Science Mission Critical Infrastructure o
Disaster Recovery and Stability
o
Recovery from Physical Attack / Failure Maintaining Science Mission Critical Infrastructure in the Face of Cyberattack
o
•
Services that Grids need from the Network o
Public Key Infrastructure example 40
Network and Middleware Needs of DOE Science August 13-15, 2002 Organized by Office of Science Mary Anne Scott, Chair Dave Bader Steve Eckstrand Marvin Frazier Dale Koelling Vicky White
Workshop Panel Chairs
• Focused on science requirements that drive o o o o
Advanced Network Infrastructure Middleware Research Network Research Network Governance Model
Ray Bair and Deb Agarwal Bill Johnston and Mike Wilde Rick Stevens Ian Foster and Dennis Gannon Linda Winkler and Brian Tierney Sandy Merola and Charlie Catlett
• The requirements for DOE science were developed by the OSC science community representing major DOE science disciplines o Climate o Magnetic Fusion Energy Sciences o Spallation Neutron Source o Chemical Sciences o Macromolecular Crystallography o Bioinformatics o High Energy Physics Available at www.es.net/#research 41
Grid Middleware Requirements (DOE Workshop)
•
A DOE workshop examined science driven requirements for network and middleware and identified twelve high priority middleware services (see www.es.net/#research)
•
Some of these services have a central management component and some do not
•
Most of the services that have central management fit the criteria for ESnet support. These include, for example o o o o o o o
Production, federated RADIUS authentication service PKI federation services Virtual Organization Management services to manage organization membership, member attributes and privileges Long-term PKI key and proxy credential management End-to-end monitoring for Grid / distributed application debugging and tuning Some form of authorization service (e.g. based on RADIUS) Knowledge management services that have the characteristics of an ESnet service are also likely to be important (future) 42
Grid Middleware Services
• ESnet provides several “science services” – services that support the practice of science
• A number of such services have an organization like ESnet as the natural provider o
ESnet is trusted, persistent, and has a large (almost comprehensive within DOE) user base
o
ESnet has the facilities to provide reliable access and high availability through assured network access to replicated services at geographically diverse locations
o
However, service must be scalable in the sense that as its user base grows, ESnet interaction with the users does not grow (otherwise not practical for a small organization like ESnet to operate) 43
Science Services: PKI Support for Grids
•
Public Key Infrastructure supports cross-site, crossorganization, and international trust relationships that permit sharing computing and data resources and other Grid services
•
DOEGrids Certification Authority service provides X.509 identity certificates to support Grid authentication provides an example of this model o
The service requires a highly trusted provider, and requires a high degree of availability
o
The service provider is a centralized agent for negotiating trust relationships, e.g. with European CAs
o
The service scales by adding site based or Virtual Organization based Registration Agents that interact directly with the users
o
See DOEGrids CA (www.doegrids.org) 44
Science Services: Public Key Infrastructure
•
DOEGrids CA policies are tailored to science Grids o
Digital identity certificates for people, hosts and services
o
Provides formal and verified trust management – an essential service for widely distributed heterogeneous collaboration, e.g. in the International High Energy Physics community
¾ This service was the basis of the first routine sharing of HEP computing resources between US and Europe
¾ Have recently added a second CA with a policy that supports secondary issuers that need to do bulk issuing of certificates with central private key management o
NERSC will auto issue certs when accounts are set up – this constitutes an acceptable identity verification
o
A variant of this will also be set up to support security domain gateways such as Kerberos – X509 – e.g. KX509 – at FNAL 45
Science Services: Public Key Infrastructure
•
The rapidly expanding customer base of this service will soon make it ESnet’s largest collaboration service by customer count Registration Authorities ANL LBNL ORNL DOESG (DOE Science Grid) ESG (Climate) FNAL PPDG (HEP) Fusion Grid iVDGL (NSF-DOE HEP collab.) NERSC PNNL
46
Grid Network Services Requirements (GGF, GHPN)
•
Grid High Performance Networking Research Group, “Networking Issues of Grid Infrastructures” (draft-ggf-ghpnnetissues-3) – what networks should provide to Grids o
High performance transport for bulk data transfer (over 1Gb/s per flow)
o
Performance controllability to provide ad hoc quality of service and traffic isolation.
o
Dynamic Network resource allocation and reservation
o
High availability when expensive computing or visualization resources have been reserved
o
Security controllability to provide a trusty and efficient communication environment when required
o
Multicast to efficiently distribute data to group of resources.
o
How to integrate wireless network and sensor networks in Grid environment 47
Transport Services
• network tools available to build services o
queue management - provide forwarding priorities different from best effort - e.g. – scavenger (discard if anything behind in the queue) – expedited forwarding (elevated priority queuing) – low latency forwarding (highest priority – ahead of all other traffic)
o
path management - tagged traffic can be managed separately from regular traffic
o
policing - limit the bandwidth of an incoming stream
48
Priority Service: Guaranteed Bandwidth bandwidth 1000 0
available for elevated priority traffic reserved for production, best effort traffic
network pipe
bandwidth management model
?
bandwidth broker
user system1
flag traffic from user system1 for expedited forwarding
site A
border router
border router user system2 site B
49
Priority Service: Guaranteed Bandwidth • What is wrong with this? (almost everything) there may be several users that want all of the premium bandwidth at the same time
the user may send data into the high priority stream at a high enough bandwidth that it interferes with production traffic (and not even know it)
?
this is at least three independent networks, and probably more a user that was a priority at site A may not be at site B
bandwidth broker
user system1
border router
site A
border router
site B
user system2 50
Priority Service: Guaranteed Bandwidth
policer
authorization
user system1
shaper
• To address all of the issues is complex
resource manager
bandwidth broker
allocation manager
site A resource manager
resource manager
user system2
site B
51
Priority Service
• So, practically, what can be done? • With available tools can provide a small number of provisioned circuits o
secure and end-to-end (system to system)
o
various Quality of Service possible, including minimum latency
o
a certain amount of route reliability (if redundant paths exist in the network)
o
end systems can manage these circuits as single high bandwidth paths or multiple lower bandwidth paths of (with application level shapers)
o
non-interfering with production traffic, so aggressive protocols may be used 52
policer
user system1
authorization
Priority Service: Guaranteed Bandwidth
bandwidth broker
resource manager
allocation will probably be relatively static and ad hoc
site A resource manager
• will probably be service level agreements among transit networks allowing for a fixed amount of priority traffic – so the resource manager does minimal checking and no authorization • will do policing, but only at the full bandwidth of the service agreement (for self protection)
resource manager
user system2
site B
53
Grid Network Services Requirements (GGF, GHPN)
•
Grid High Performance Networking Research Group, “Networking Issues of Grid Infrastructures” (draft-ggf-ghpnnetissues-3) – what networks should provide to Grids o
High performance transport for bulk data transfer (over 1Gb/s per flow)
o
Performance controllability to provide ad hoc quality of service and traffic isolation.
o
Dynamic Network resource allocation and reservation
o
High availability when expensive computing or visualization resources have been reserved
o
Security controllability to provide a trusted and efficient communication environment when required
o
Multicast to efficiently distribute data to group of resources.
o
Integrated wireless network and sensor networks in Grid environment 54
High Throughput Requireme nts
1) High average throughput 2) Advanced protocol capabilities available and usable at the end-systems 3) Lack of use of QoS parameters
Current issues
1) Low average throughput 2) Semantic gap between socket buffer interface and the protocol capabilities of TCP
Analyzed reasons
1a) End system bottleneck, 1b) Protocol misconfigured, 1c) Inefficient Protocol 1d) Mixing of congestion control and error recovery 2a) TCP connection Set up: Blocking operations vs asynchronous 2b)Window scale option not accessible through the API
Available solutions
1a) Multiple TCP sessions 1b) Larger MTU 1c) ECN
Proposed alternatives
1) Alternatives to TCP (see DT-RG survey document) 2) OS by-pass and protocol off-loading 3) Overlays 4) End to end optical paths 55
A New Architecture
• The essential requirements cannot be met with the current, telecom provided, hub and spoke architecture of ESnet o (CHI) Chicag
DOE sites
New York (AOA)
ESnet Core/Backbone
Washington, DC (DC)
Sunnyvale (SNV) El Paso (ELP)
Atlanta (ATL)
• The core ring has good capacity and resiliency against single point failures, but the point-topoint tail circuits are neither reliable nor scalable to the required bandwidth 56
A New Architecture
• A second backbone ring will multiply connect the MAN rings to protect against hub failure AsiaPacific
Europe
o( Chicag
CHI)
DOE sites
Sunnyvale (SNV)
New York (AOA)
ESnet Core/Backbone
Washington, DC (DC) Atlanta (ATL)
El Paso (ELP)
• All OSC Labs will be able to participate in some variation of this new architecture in order to gain highly reliable and high capacity network access
57
Conclusions
• ESnet is an infrastructure that is critical to DOE’s science mission and that serves all of DOE
• Focused on the Office of Science Labs • ESnet is working on providing the DOE mission science networking requirements with several new initiatives and a new architecture
• QoS is hard – but we have enough experience to do pilot studies (which ESnet is just about to start)
• Middleware services for large numbers of users are hard – but they can be provided if careful attention is paid to scaling 58
