Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
Dr.Stefan Radtke
Dr. Stefan Radtke
Europe's largest Internet Service Platform powered by RS/6000SP Systems -A Technical Overview -
Dr.Stefan Radtke
TopicsInternet Service Platform - Introduction
Customer Groups of the PlatformCore Components of DT's Platform developed by IBM
Billing Broker ComponentBilling Broker Component as a Core ServiceCharacteristics of the Dial In ServiceUser Dial In ScenarioHow high availability is implementedPhysical Configurations (SP,IP-Net, Storage)Distributed Service to support 50 Mio customers
Web Cache ServiceWhy are Telco's invest in caching Object size distribution vs. bandwidth requirementsHow does the Websphere Edge Server Cluster workHardware platformGrowth of IP trafficStatistic and performance tools
Miscellaneous (...only when sufficient time left)Caching of Streaming Data (RTSP)Software distributionPhysical RS/6000SP configurations
Dr.Stefan Radtke
Customer Groups of the ISP
IP Platform
Open Communication
Closed Communication
Virtual Internet Provider
- T-Online- ....- ....
Virtual Private Networks
- Corporate Networks- ...
Realization ofVPNs
based on the samearchitecture
Today more than 150Virtual Internet Provider
are served by the platformwith about 10 Mio. users
Authentication, Authorization, Accounting
Content Provider
- Email- Web Sites- Video On
Demand - ASP...
Content Provider offer Services
to known Users
Services
Dr.Stefan Radtke
Core Components of the ISPBilling Broker Component(Protocol = RADIUS, Remote Access Dial In User Service) Distributed and higly available application for AAA-Services: - Authentication - Authorization - Accounting Handles about 10 Million users today but designed for 50 Mil.
Web Caching Scalable solution for - Saving bandwith - Faster Web access - Saving peering costs Serves more than 500.000.000 HTTP Requests per day to endusers
- Video on Demand Video platform for vendor independent video formats
Dr.Stefan Radtke
TopicsInternet Service Platform - Introduction
Customer Groups of the PlatformCore Services of DT's Platform developed by IBM
Billing Broker ComponentBilling Broker Component as a Core ServiceCharacteristics of the Dial In ServiceUser Dial In ScenarioHow high availability is implementedPhysical Configurations (SP,IP-Net, Storage)Distributed Service to support 50 Mio customers
Web Cache ServiceWhy are Telco's invest in caching Object size distribution vs. bandwidth requirementsHow does the Websphere Edge Server Cluster workHardware platformGrowth of IP trafficStatistic and performance tools
Miscellaneous (...only when sufficient time left)Caching of Streaming Data (RTSP)Software distribution
Physical RS/6000SP configurations
Dr.Stefan Radtke
Central Component of the ISP
Billing Broker Component
ISDN
Network Access Server
NAS
ADSL
RADIUS
E-MailServer
POP,IMAP
RADIUS *)
Video Server
RTSP
Video Provider
CustomerCareCustomerCare
Billing
VIP / VPN / ASP
CustomerCareCustomerCare
BillingCustomerCareCustomerCare
Billingftp
Analog
ISP Dial In ServiceASP E-Mail
ASP Video on Demand
ASP 1
ASP 3ASP 2
e.g. Media Hosting,Web Hosting
RADIUS *)
RADIUS *)
Mobile
MQSeries RADIUS
Dr.Stefan Radtke
Characteristics of Dial In Service Distributed
POPs are georaphically distributed 13 POPs with RADIUS-Proxy and RADIUS-LocalServer Instances Central Database with Millions of Data Records are replicated to 13 POPs Distributed Database
Scaleable Distributed Architecture is designed for more than 50 Million Users Distributed Database & Web-Cache 3 Tier RADIUS Architecture
Highly Available Service are all highly available through Application redundance All important infrastructure components are redundant
Driven by IBM TechnologyRADIUS-Proxy Server and Local Server running on AIX, PSSP, RS/6000SPIBM Middleware for highly available Services: DB2 UDB, HACMPIBM Storage Technology for performance and availablilityCentral RADIUS Server running on zSeries-OS/390 in parallel Sysplex with DB2
Intregration of 3rd Party Components Cisco Routers Network Access Routers from Lucent, Shiva, Accent, Siemens
Dr.Stefan Radtke
User Dial In Scenario (1)
ISDN
Network Access Server
NAS
ADSL
Internet
RADIUS-Proxy
Local Server/DB
RADIUS-ProxyRADIUS-ProxyIP-Addresses
User Data
Firewall
Central Server
Caching Proxy Cluster
httpd
NASUser Data
POP (13x)
1. User Request to NAS2. NAS talks to RD-Proxy3. Proxy asks Local Server4. Auth.accept to Proxy5. Proxy allocates IP-Address6. Auth. accept Proxy to NAS7. Accounting start to Central- Server8. User is connected to Network with dynamic IP-Address
Analog
Dr.Stefan Radtke
ISDN
Network Access Server
NAS
ADSL
Internet
Central Server
User Data
PO
P (
13x)
HA Implementation on RS/6000SP
RS
/600
0SP
zSeries
Central Server Central Server Central Server
Parallel Sysplex
HACMPCluster
Radius-Proxy
7133-600
Mirror
Radius-Proxy(Backup)
7133-600Radius-Svr Radius-DB
(Backup)
Firewall Firewall
7133-600Radius-Proxy Radius-Proxy
(Backup)
Analog
Dr.Stefan Radtke
TopicsInternet Service Platform - Introduction
Customer Groups of the PlatformCore Services of DT's Platform developed by IBM
Billing Broker ComponentBilling Broker Component as a Core ServiceCharacteristics of the Dial In ServiceUser Dial In ScenarioHow high availability is implementedPhysical Configurations (SP,IP-Net, Storage)Distributed Service to support 50 Mio customers
Web Cache ServiceWhy are Telco's invest in caching Object size distribution vs. bandwidth requirementsHow does the Websphere Edge Server Cluster workHardware platformGrowth of IP trafficStatistic and performance tools
Miscellaneous (...only when sufficient time left)Caching of Streaming Data (RTSP)Software Distribution
Physical RS/6000SP configurations
Dr.Stefan Radtke
Caching Proxy Cluster on RS/6000SP
1. Reduce Response Time for Customers
2. Reduce required network bandwidth
3. Reduce costs for IP traffice over peering points
4. Reduce load on overloaded web servers
Why are Telco's investing in Web Cache ?
Dr.Stefan Radtke
HTTP Traffic growth
Today the WES cluster serves about 500.000.000 HTTP Requests per day
Reqeuests/Day
050000000
100000000150000000200000000250000000300000000350000000400000000450000000
Jan
98
Apr
98
Jun
98
Aug
98
Oct
98
Jan
99
Mar
99
May
99
Jul 9
9
Oct
99
Dec
99
Feb
00
Apr
00
Jun
00
Sep
00
Nov
00
Feb
01
Dr.Stefan Radtke
HTTP Traffic growth
... and about 4 TB of data through the caching proxy clusters
Bytes/Day
0500000000000
100000000000015000000000002000000000000250000000000030000000000003500000000000400000000000045000000000005000000000000
Jan
98
Apr
98
Jun
98
Sep
98
Nov
98
Jan
99
Apr
99
Jun
99
Sep
99
Nov
99
Jan
00
Apr
00
Jun
00
Sep
00
Nov
00
Feb
01
Dr.Stefan Radtke
Object Size vs. Bandwidth Consumption
<1K
1-10
K
10-1
00K
100K
- 1M
1 - 1
0M
10 -
100M
0102030405060
Percent of Repeat Objects
<1K
1-10
K
10-1
00K
100K
- 1M
1 - 1
0M
10 -
100M
0
10
20
30
40
Percent of Repeat Traffic ... those objects consume over 50% of network bandwidth!
Very few (<1%) object requests are for objects larger than 1 MB...
Dr.Stefan Radtke
WWW-Performance for 8 Million Users
2 Tier Load Balancing Telephone Network routes call to nearest POP Load Balancing within the POP by Network Dispachtcher
'Local' DNS Server on each Cluster Node DNS Server on ND-Nodes answer user queries DNS Server on WTE-Nodes answer WTE queries
Caching Proxy Cluster with dynamic Load Balancing Using IBM Interactive Session Support (ISS) Takes dynamically into account different CPU capabilities Streaming formats like Real Video/Audio can be integrated
Dr.Stefan Radtke
2 Tier Load Balancing
ND
TelephoneNetwork
NAS
NAS
NAS
NAS
ND
ND
ND
Internet
Network Dispatcherand 'local' DNS Cache Server
User connections are routed to the geographically nearestPOP & Proxy Cluster.
Dr.Stefan Radtke
IBM Websphere Edge Server
Case 1: Requested data already in local cache
Case 2: Requested data saved in neighbour cache
Case 3: Requested data not yet saved within cluster
Caching Web with Remote Cache Access
Dr.Stefan Radtke
Case 1: Requested Data is already in Cache
WTE1
WTE2
WTE3
ND
ND
WTE4
WTE5
WTE6
Internet
www-proxy.dtag.deGEThttp://www.abc.com
1
2
4 - Direkte Antwort
WTE7
WTE8
160MB/s
160MB/s
160MB/s
160MB/s
160MB/s
160MB/s
160MB/s
160MB/s
3
Dr.Stefan Radtke
WTE1
WTE2
WTE3
ND
ND
WTE4
WTE5
WTE6
Internet
GEThttp://www.abc.com
1
2
6 - Direkte Antwort
WTE7
WTE8
160MB/s
160MB/s
160MB/s
160MB/s
160MB/s
160MB/s
160MB/s
160MB/s
4
53
3,5 = HTTP
Case 2: Data resides in Neighbour Cache
Dr.Stefan Radtke
Case 3: Data not yet available in Cache
WTE1
WTE2
WTE3
ND
ND
WTE4
WTE5
WTE6
Internet
GEThttp://www.abc.com
1
2
7 - Direkte Antwort
WTE7
WTE8
160MB/s
160MB/s
160MB/s
160MB/s
160MB/s
160MB/s
160MB/s
160MB/s
5
63
4
Dr.Stefan Radtke
Caching Proxy Cluster: Physical Configuration Deutsche Telekom AG runs about 50 Caching Proxy Clusters based on RS/6000SP technology with about 400 nodes.
A1A2B1B2
A1A2B1B2
B2B1A2A1
B2B1A2A1
B2B1A2A1 7133-600
4
5
8
9
8 123456
9 10 1112 13 14 15 16
1
12
13
16
7
9,1 9,19,19,1
9,1 9,19,19,1
9,1 9,19,19,1
9,1 9,19,19,1
B2B1A2A1
B2B1A2A1
B2B1A2A1
B2B1A2A1
B2B1A2A1
7133-600
4
5
8
9
8 123456
9 10 1112 13 14 15 16
1
12
13
16
7
9,1 9,19,19,1
9,1 9,19,19,1
9,1 9,19,19,1
9,1 9,19,19,1
WTE + ND
WTE + ND
WTE
WTE
WTE
WTE
WTE
WTE
- 8 x SP Nodes- each node with 1 GB RAM- each node has 4x 9,1 GB SSA Disks - CacheSize per cluster = 8x 36 GB = 288 GB- ND and Backup-ND are running co-located
Dr.Stefan Radtke
Dynamic Load Balancing with ISSDifferent nodes have different performance andcan handle different number of requests/time.
Typ T-Web-Proxy-Prod. SpecWeb99-----------------------------------------------------------------------------------------------------------Power2SC 160MHz 1 CPU: ca. 4.500 req/min.PowerPC 332Mhz 2 CPUs ca. 6.000 req/min.Power3 200Mhz 2 CPUs 14.000 req/min. 440Power3-II 370 Mhz 4 CPUs ~ 40.000 req/min*) 1359 *) estimated
Number of HTTP-Reqeusts per Minute
The test nodes were equiped with - 1 GB RAM - 4 oder 5x 9.1 GB SSA Disks
DynamicLoad Balancingis required !!
Dr.Stefan Radtke
eND Cluster with ISS
ISS Agent
ISS Agent
ISS Agent
ISS Agent
ISS Agent
ISS Agent
ISS Agent
ISS Monitor
ISS Agent
Dispatcher
Load Information
Requests
Customized ND-Proportions 4 % Active Connections 3 % New Connections 3 % Advisor90 % System load
advisoradvisor
advisor
manager
executor
ndcontrol
ndserver
Manager: sets weights based on - internal counters in executor - feedback from servers/advisors - system monitoring, e.g. ISSAdvisors: query servers and calls Manager for setting weights.
- exist for HTTP, SSL, POP,SMTP,NNTP,IMAP,... - runs optional (but recommended !!)Executor: does the load balancing. - mandatory
Dr.Stefan Radtke
Customized Dispatcher Weights
Customized ND-Proportions: 4 % Active Connections 3 % New Connections 3 % Advisor90 % System load (ISS)
Default ND-Proportions: 90 % Active Connections 5 % New Connections 5 % Advisor 0 % System load
We changed the default weight proportions since we are usingdifferent nodes. The system load*) takes automatically into account thedifferent I/O and CPU capabilities.
*) Load Average is the number of processes which were ready to run (though not waiting for I/O) but waiting to be dispatched to CPU ... during the last 1, 10, 15 minutes.
Dr.Stefan Radtke
Average Response Times
Average response time over 24 hours for 20 http GET requests
Result: Responses over Caching-Proxies are about 2 times faster than direct requests !!
24 Stunden Durchschnitt
0
5
10
15
20
25
30
35
Proxy Direkt
Seku
nden
für 2
0 R
equ
ests
0
5
10
15
20
25
30
35
Seku
nden
für 2
0 R
equ
ests
Resonse Timefor 20 http get requests:
over Proxy: 15 sec.direct: 32 sec.
Dr.Stefan Radtke
Statistics
Detailed statistics are available for each cluster and each node. - Special plugin developed (performance) - Available in table- and grafic format through Web interface.
Dr.Stefan Radtke
TopicsInternet Service Platform - Introduction
Customer Groups of the PlatformCore Services of DT's Platform developed by IBM
Billing Broker ComponentBilling Broker Component as a Core ServiceCharacteristics of the Dial In ServiceUser Dial In ScenarioHow high availability is implementedPhysical Configurations (SP,IP-Net, Storage)Distributed Service to support 50 Mio customers
Web Cache ServiceWhy are Telco's invest in caching Object size distribution vs. bandwidth requirementsHow does the Websphere Edge Server Cluster workHardware platformGrowth of IP trafficStatistic and performance tools
Miscellaneous (...only when sufficient time left)Caching of Streaming Data (RTSP)Software distribution
Physical RS/6000SP configurations
Just for backup
Dr.Stefan Radtke
Caching of Streaming formats
RTSP = Real Time Streaming ProtocolThe Real Time Streaming Protocol is defined in RFC 2326. It is an Internet standard protocol for controlling data streams. It does not include technology for delivering streams
RTSP Redirector im WTEThe RTSP redirection feature allows Web Traffic Express to redirect requests for any streaming media sessions controlled by RTSP. These include the following types of media:
RealNetworks recorded audio RealNetworks recorded video RealNetworks live streams (audio and video) Microsoft Media Player files Apple Quicktime media files
Any player that can be configured to contact a proxy server on its RTSP port (typically 554) can use this framework in WTE to have its requests handled by the RTSP redirector.
Dr.Stefan Radtke
RTSP Redirect (1/2)
WTE
ND
ND
WTE
WTE
WTE
1
2
RS/6000SP
GET http://..../wantit.ra&proto=rtsp HTTP/1.0
HTTP/1.1 200 OK.Server: Lightweb/0.00 (Linux)Content-type: audio/x-pn-realaudio<CR><CR>rtsp://g2...com/.../wantit.ra
3
4
5
Cache f. Streaming Daten
Streaming Proxies Real-Networks,Microsoft,Quicktime,...
Dr.Stefan Radtke
WTE
ND
ND
WTE
WTE
WTE
1
2
RS/6000SP
Streaming Proxies Real-Networks,Microsoft,Quicktime,...
5
rtsp://g2...com/.../wantit.ra
3
46
Cache f. Streaming Daten
RTSP Redirect (2/2)
3.) RTSP/1.0 305 UseProxy Location: rtsp://193.158.128.70:554/
4) OPTIONS rtsp://g2-rbn-nfs06.rbn.com:554 RTSP/1.0 User-Agent: RealMedia Player Version 6.0.7.19
5) Autorisierungs Session
6) Auslieferung Video/Audio aus Cache
Dr.Stefan Radtke
RTSP Trace between (1/3) Client->WTE->Real-Proxy->RealServer
Client > WTE:80GET http://play.rbn.com/?url=gmusic/gmusic/g2demand/peeps/bboys/wantit.ra&proto=rtsp HTTP/1.0Referer: http://www.peeps.com/bboys/media.htmlProxy-Connection: Keep-AliveUser-Agent: Mozilla/4.5 [en] (Win98; I)Host: play.rbn.com..Accept: image/gif, image/x-xbitmap, image/jpeg, image/pjpeg, image/png, */*Accept-Encoding: gzip..Accept-Language: enAccept-Charset: iso-8859-1,*,utf-8<CR><CR>
WTE > Real-Server:80GET /?url=gmusi/gmusic/g2deman/peeps/bboys/watit.ra&proto=rtp HTTP/1.1Hos: play.rbn.com.Connection: Kee-Alive, TERefrer: http://www.peeps.com/bboys/media.htmlAccept: image/gif, image/x-xbitmap, image/jpeg, image/pjpeg, image/png, */*Accept-Encoding: gzipAccept-Language:enAccept-Charset: iso-8859-1,*,utf-8TE: chunkedVia: HTTP/1.0 spcss47. (IBM-PROXY-WTE)Client-IP: 193.158.128.84User-Agent: Mozilla/4.5[en] (Win98; I) <CR><CR>
Dr.Stefan Radtke
Trace (2/3)Real-Server:80 > WTEHTTP/1.1 200 OK.Server: Lightweb/0.00 (Linux)Content-type: audio/x-pn-realaudiortsp://g2-rbn-nfs06.rbn.com/gmusic/gmusic/g2demand/peeps/bboys/wantit.ra
WTE:80 > ClientHTTP/1.1 200 OKConnection: closeProxy-Connection: closeVia: HTTP/1.1 spcss47 (IBM-PROXY-WTE)Server: Lightweb/0.00 (Linux)Content-type: audio/x-pn-realaudio<CR> <CR>
WTE :80> Clientrtsp://g2-rbn-nfs06.rbn.com/gmusic/gmusic/g2demand/peeps/bboys/wantit.ra.
Client > WTE.554 (RTSP-Redirector)OPTIONS rtsp://g2-rbn-nfs06.rbn.com:554 RTSP/1.0CSeq: 1User-Agent: RealMediaPlayer Version 6.0.7.194 (win32)ClientChallenge: ab8bd4308b03704d9773d0eb3890dfb6PlayerStarttime: [11/07/2000:20:18:27 00:00]CompanyID: W9LrkNVmrQ4si74ItX7pcQ==GUID: caafabe0-5588-11d3-a6fd-0000e8557559RegionData:ClientID: Win98_4.10_6.0.6.99_play32_MF60_en-US_586<CR> <CR> WTE.554 (RTSP-Redirector) > Client RTSP/1.0 305 UseProxyCSeq: 1L ti t //193 158 128 70 554/
Dr.Stefan Radtke
Trace (3/3)
WTE.554 (RTSP-Redirector) > Client RTSP/1.0 305 UseProxyCSeq: 1Location: rtsp://193.158.128.70:554/Date: TueJul 11 20:19:482000Server: Web Traffic Express/Edge Server RTSP Redirector 0.1<CR> <CR> Client > Real-Proxy-Server.rtsp: OPTIONS rtsp://g2-rbn-nfs06.rbn.com:554 RTSP/1.0CSeq: 1User-Agent: RealMedia Player Version 6.0.7.19
Real-Proxy-Server.3253 > Real-Origin-Server .rtspOPTIONS rtsp://g2-rbn-nfs06.rbn.com:554 RTSP/1.0CSeq: 1User-Agent: RealMedia Player Vers
Dr.Stefan Radtke
3-Tier Software DistributionSoftware Packages are replicated (rdist) over night to about50 Control Workstations (with ssh).
Only one repository must bemaintained.
Installation to nodes fromCWS does not consumeWAN capacity.
Dr.Stefan Radtke
Physical Configuration on RS/6000SP
RARs
neue RARs
Cisco Router
LAN-Switch 1
LAN-Switch 2
Control-Ethernet (en0)(all Nodes)
VPN-Netz (en5,en6)
AGB-Netz (en3,en4)
100/100 Mbit Ethernet Adapter
SSA-Adapter
PCI-Steckplatz leer
HACMP-Stby-Net (en2)
HACMP-Service-Net (en1)
LAN-Switch 3
Ethernet-Hub
1
3
5
7
9
11
13
15
IP-Net
Dr.Stefan Radtke
1
3
5
7
9
11
13
15
B2B1A2A1
B2B1A2A1
B2B1A2A1
B2B1A2A1
B2B1A2A1
B2B1A2A1
B2B1A2A1
B2B1A2A1
B2B1A2A1
B2B1A2A1
B2B1A2A1
4
5
8
9
1
12
13
16radiusdb3
8 45 13
9 10 12 13 15 16
3 2
14
7 6
11
4
5
8
9
1
12
13
16
rad1
vg
rad2
vg
rad4
vg
rad3
vg
8 45 13
9 10 12 13 15 16
3 2
14
7 6
11
radiusdb1
4
5
8
9
1
12
13
168 45 13
9 12 13 16
3 2
14
7 6
1110 15
B2B1A2A1
radi
usdb
ls1vg
ls2vgls2vg
Physical Configuration on RS/6000SPStorage (SSA)