Transcript Imperial

1
IC HEP Computing Review
Geoff. Fayers
[email protected]
2
Current Experiment
Programme






Aleph
BaBar
CMS
Dark Matter (Boulby)
Diamond-based
applied research
D0



LHCb
Silicon-based detector
development (esp.
CMS)
Zeus
3
Current Computing Hardware








DEC4610 AXP
5 DEC AXP W/stns
4 more?
2 SUN Enterprise Svrs
2 SUN W/stns
1 NT Server+BDC
~37 PCs
~18 X-terms






Ageing Macs + some
new G3+later
2 RIOS PPCs (lynx)
~150 assigned IP addrs
~1.25Tb storage excl.
some local PC disks
Printers mostly HP
1 VAXStn 4000/60
Current Software







DUX 4.0D on all AXP
Solaris 2.6 on
Enterprises (BaBar)
Solaris 2.7 otherwise
SAMBA
NT4 + Office97/2000
=eXceed at desktop
NT4 Server
Linux5.2 (D0 FNAL
variety)







AFS on all UNIX
Europractice Selection
incl. Synopsys,
HSPICE, Xilinx, et al.
AutoCAD2000 ++
Ranger PCB design
MS, Adobe packages
for NT
(Optivity)
Dr. Solomon’s AVtk
4
Some Features
5
Homogeneous systems, NIS, NIS+, DNS
 Automount distributed file systems
 Quotas on home dirs.
 User-based scratch areas, 7-day files’ lifetime
 User-based stage areas
 Per-experiment data areas on rwx rwx r-x permissions
 Recent news headlines at login
 Optional news via e/mail
 Process priority monitor

Concerns - 1
Non-convergence of Linux versions for LHC, US?
 Potential fragmentation of resources
 Non-scalability of current storage deployment
 Management effort for NT: SMS, other?
 No NT YP map capability (except Solaris6/NT4)
 Vagaries of A/V software
 Possible management of all http traffic by ICCS
 Costs of MS licences and network management tools
 Will 64-bit Linux (e.g IA64) supplant PC arch.?
 Possible withdrawal of ULVC Europractice support

6
7
Concerns - 2
Poor US connectivity esp. to FNAL for D0
 Network management tools?
 Short investment cycles for PCs
 Metaframe expensive thin client solution
 Uncertain future PCI replacement
 Security, security, security, ……...

Networking

< Summer 1999:
–
–
–
–
–
= 10Base5, 10Base2
16-port 10BaseT Switch, HDx only
NFS traffic via 2nd NICs through switch
5/10/20/80m non-standard custom braided fly leads only
Limited division of 10Base2 collision domain to ease
congestion
– Tolerable response (mostly)
– WAN via DECnis600 router: 10Base5 to AGS4+ on
Campus FDDI (IP routing only)
8
9
Networking Upgrade
Upgrade Summer 1999
 Decisions:

–
–
–
–
–
Expected lifetime
Likely technologies and services required
Density of service ports
Future campus connectivity
Supporting hardware
10
Expected Lifetime
Existing 10Base5/2 10 years old  new
system at least 10 years
 DECnis600 ~6 years  new network
hardware at least 5 years

Future Technologies & Services - 1

Summer 1999:
– 803.1d bridging standard updated and fully ratified as
802.1D-1998 (all 802.* technologies embraced)
– GBE Adaptors at commodity prices, but...
PCs too slow
 PCI-X interim solution
 GBE as 1000-BaseSX and 1000-BaseLX only
 GBE over Class D/E UTP expected Mar. 2000
 Proprietary Class E solutions emerging for 4 x 250MHz
 Class F as 4 x 300MHz on horizon (but irrelevant for HEP?)

11
Future Technologies & Services - 2
IEEE HSSG already convened (appendix)
 LANe still complicated
 DWDM very likely in SJ4 - ATM will wane

12
The Other Issues









Current aggregated b/w > 3.7 GB/sec, so Catalyst 5000 and
some competitors useless (received wisdom: 2 x aggregate)
Increasing use of laptops
Outlet density to ISO standards - high
No ATM adaptors for CoreBuilder 3500’s, so…...
ICCS ATM spine strategy ditched…….
Campus CellPlex7000’s dumped in favour of 9300’s (GBE)
Stuck with CoreBuilder 3500 100BaseFX uplinks (poor
performance, 802.3q implementation lagging)
Possible need for ISDN
DECnis600 replacement routing capability
13
Structured Cabling Decision
Full Cat5e compatibility and compliance mandatory
 Mixture of Class D and Class E proprietary UTP
 MMF and SMF unterminated in corridor ceiling voids
 Insist on Class E performance testing to 250MHz
 Insist on full compliance to draft Class E standard
 Cat6 Guarantee underpinned by Manufacturer
Indemnity
 Only tied Certified Installers considered
 Leverage early take-up and prestige potential 

14
Implementation
15
8 ‘Cat6’ passive component manufacturers evaluated
 Only one testing fully in 200 - 300 MHz envelope
 Got flood Cat6 at 10% more than Cat5
 17.5GB/sec non-blocking Extreme Summit48, level 2
and level 3 switching, 4-levels QoS (upgradable)
 2 x 1000BaseSX ports, optional 1000BaseLX trcvrs
 HP ProCurve 2424M 10/100 switches
 HP ProCurve 24 10/100 hubs
 100BaseFX uplinks to Campus spine

16
Problems
Completed Dec 1999
 Test results presented on CD OK but….
 Manufacturing fault on some ports - resolution in
hand…...
 HP 2424M fault…….HP to replace (lifetime on-site
warranty)

Future
RAID-based storage, possibly home-grown solution
 Non-proprietary SAN based on dual cpu 64-bit
server, GBE to Extreme
 Linux Farm(s) with 100BaseT cpus
 DHCP
 WAN connectivity resilience via FOIL?
 Possible Campus Spine bypass?
 Managed bandwidth pilot (BaBar)
 H.323/320 VC

17