Soluzioni HW per il Tier 1 al CNAFSoluzioni HW per il Tier 1 al CNAF

Luca dell’Agnello Stefano Zani

(INFN – CNAF, Italy)

III CCR Workshop May 24-27 2004

Tier1

INFN computing facility for HEP community Ending prototype phase last year, now fully operational Location: INFN-CNAF, Bologna (Italy)

o One of the main nodes on GARR network Personnel: ~ 10 FTE’s

o ~ 3 FTE's dedicated to experiments Multi-experiment

LHC experiments(Alice, Atlas, CMS, LHCb), Virgo, CDF, BABAR, AMS, MAGIC, ...

Resources dynamically assigned to experiments according to their needs

50% of the Italian resource for LCG Participation to experiments data challenge Integrated with Italian Grid Resources accessible also in traditional way

Logistics

Recently moved to a new location (last January) Hall in the basement (-2nd floor) ~ 1000 m2 of total space

o Computing Nodes o Storage Deviceso Electric Power System (UPS)o Cooling and Air conditioning systemo Garr GPop

Easily accessible with lorries from the road Not suitable for office use (remote control needed)

Electric Power

Electric Power Generator 1250 KVA (~ 1000 KW)

up to 160 racks Uninterruptible Power Supply (UPS)

Located into a separate room (conditioned and ventilated) 800 KVA (~ 640 KW)

380 V three-phase distributed to all racks (Blindo) Rack power controls output 3 independent 220 V lines for

computers Rack power controls sustain burden up to 16 or 32 A

o 32 A power controls needed for Xeon 36 bi-processors racks 3 APC power distribution modules (24 outlets each)

o Completely programmable (allows gradual servers switching on)o Remotely manageable via web

380 V three-phase for other devices (tape libraries, air conditioning, etc…)

Cooling & Air Conditioning

RLS (Airwell) on the roof ~ 700 KW Water cooling Need “booster pump” (20 mts T1 roof) Noise insulation

1 Air Conditioning Unit (uses 20% of RLS refreshing power and controls humidity)

12 Local Cooling Systems (Hiross) in the computing room

WN typical Rack Composition

• Power Controls (3U)• 1 network switch (1-2U)

– 48 FE copper interfaces

– 2 GE fiber uplinks

• 34-36 1U WNs– Connected to network

switch via FE

– Connected to KVM system

Remote console control

Paragon UTM8 (Raritan) 8 Analog (UTP/Fiber) output connections Supports up to 32 daisy chains of 40 nodes (UKVMSPD

modules needed) Costs: 6 KEuro + 125 Euro/server (UKVMSPD module) IP-reach (expansion to support IP transport) evaluted but not

used Autoview 2000R (Avocent)

1 Analog + 2 Digital (IP transport) output connections Supports connections up to 16 nodes

o Optional expansion to 16x8 nodes Compatible with Paragon (“gateway” to IP)

Evaluating Cyclades Alterpath KVM via serial line (cheaper)

Networking (1)

Main Network infrastructure based on optical fibres (~ 20 Km)

To ease adoption of new (High Performances) transmission technologies

To insure a better electrical insulation on long distances Local (Rack wide) links with UTP (copper) cables

LAN has a “classical” star topology GE core switch (Enterasys ER16) NEW core switch is going to be shipped (Next july)

o 120 Gbit Fiber (Scale up to 480 ports)o 12 10 Gbit Ethernet (Scale up to max 48 ports)

Farms up-link via GE trunk (Channel) to core switch Disk Servers directly connected to GE switch (mainly

fibre)

Networking (2)

WN's connected via FE to rack switch (1 switch per rack)

Not a single brand for switches (as for wn's)o 3 Extreme Summit 48 FE + 2 GE portso 3 3550 Cisco 48 FE + 2 GE portso 8 Enterasys 48 FE 2GE portso 7 switch Summit400 48 GE copper + 2 GE ports + (2x10Gb

ready) Homogeneous characteristics

o 48 Copper Ethernet portso Support of main standards (e.g. 802.1q)o 2 Gigabit up-links (optical fibers) to core switch

CNAF interconnected to GARR-G backbone at 1 Gbps. Giga-PoP co-located 2 x 1 Gbps test links to CERN, Karlsruhe

FarmSW3(IBM)

NA

S4

FarmSWG1

SSR8600

Bo 12KGP

FarmSW1

FarmSW2(Dell)

LHCBSW1

NA

S2

NA

S3

S.Zani

FarmSW4(IBM3)Catalyst3550

FarmSW5(3Com)

DE

LL

A

XU

S

SA

N

Disk Server

F.C.

F.C.

F.C.

F.C.

FarmSW9

FarmSW12131.154.99.121

FarmSW6

FarmSW7

FarmSW8 FarmSW10

FarmSW11

FarmSWG2

ST

K

F.C.

1st Floor

Internal services

T1

Babar SW

NA

S1

Network Configuration

Info

rtrend

F.C.

L2 Configuration

Each Experiment has its own VLAN Solution adopted for complete granularity

Port based VLAN VLAN identifiers are propagated across switches

(802.1q) Avoid recabling (or physical moving) of machines to

change farm topology Level 2 isolation of farms Possibility to define multi-tag (Trunk) ports (for

servers)

Power Switches

• 2 models used at Tier1:• “Old” APC MasterSwitch

Control Unit AP9224 controlling 3x8 outlets 9222 PDU from 1 Ethernet

• “New” APC PDU Control Unit AP7951 controlling 24 outlets from 1 Ethernet

• “zero” Rack Unit (vertical mount)

• Access to the configuration/control menu via serial/telnet/web/snmp

• 1 Dedicated machine running APC Infrastruxure Manager Software (in progress)

See also: http://www.cnaf.infn.it/cnafdoc/CD0044.doc

Remote Power Distribution Unit

Screenshot of APC Infrastruxure Manager Software with the status of all TIER1 PDU

                                                 

Computing units

~ 400 1U rack-mountable Intel dual processor servers

800 MHz – 2.4 GHz ~ 240 wn’s (~ 480 CPU’s) available for LCG

To be shipped June 2004: 32 1U bi-processors Pentium 2.4 GHz 350 1U bi-processors Pentium IV 3.06 GHz

o 2 x 120 GB HDso 4 GB RAM o 2159 euro each

Tendering: HPC farm with MPI

o Servers interconnected via Infiniband Opteron farm (near future)

o To allow experiments to test their software on AMD architecture

Storage Resources

~50 TB RAW Disk Space ON LINE. NAS

o NAS1+NAS4 (3Ware low cost) Tot 4.2 TBo NAS2+NAS3 (Procom) Tot 13.2 TB

SAN o Dell Powervault 660f Tot 7 TBo Axus (Brownie) Tot 2 TBo STK Bladestore Tot 9 TBo Infortrend ES A16F-R Tot 12 TBo IBM Fast-T 900 (in few weeks) Tot 150 TB

See also:http://www.lnf.infn.it/sis/preprint/pdf

/INFN-TC-03-19.pdf

STORAGE resource

CLIENT SIDE

WAN or TIER1 LAN

PROCOM NAS2Nas2.cnaf.infn.it8100 GbyteVIRGO ATLAS

Fileserver CMS diskserv-cms-1

PROCOM NAS3Nas3.cnaf.infn.it4700 GbyteALICE ATLAS

IDE NAS1,NAS4Nas4.cnaf.infn.it1800+2000 GbyteCDF LHCB

AXUS BROWIECirca 2200 GByte 2 FC interface

DELL POWERVAULT7100 GByte2 FC interface

FAIL-OVERsupport

Gadzoox SlingshotFC Switch 18 port

RAIDTEC1800 Gbyte2 SCSI interfaces

CASTORServer+staging

STK180 with 100 LTO (10Tbyte Native)

Fileserver Fcds2

Alias diskserv-ams-1 diskserv-atlas-1

STK BladeStoreCirca 10000 GByte 4 FC interface

STK L5500 robot (max 5000) 6 LTO-2

InfortrendES A16F-R12 TB

Storage management and access (1)

Tier1 storage resources accessible as classical storage or via grid

Non grid disk storage accessible via NFS Generic WN’s also have AFS client NFS mount volumes configured via autofs and

ldap unique configuration repository eases maintenance in progress: integration of ldap configuration with Tier1

db data Scalability issues with NFS

Experienced stalled mount points Recent nfs versions use synchronous export: needed to

revert to async and use reduced rsize and wsize to avoid huge amount of retransmissions

Storage management and access (2)

Part of disk storage used as front-end to CASTOR Balance between disk and CASTOR according to

experiments needs 1 stager for each experiment (installation in

progress) CASTOR accessible both directly or via grid

CASTOR SE available ALICE Data Challenge used CASTOR architecture

Feedback to CASTOR team Need optimization for file restaging

Tier1 Database

Resource database and management interface Postgres database as back end Web interface (apache+mod_ssl+php) Hw servers characteristics Sw servers configuration Servers allocation

Possible direct access to db for some applications Monitoring system Nagios

Interface to configure switches and interoperate with installation system.

Vlan tags dns dhcp

Installation issues

Centralized installation systemLCFG (EDG WP4)Integration with a central Tier1 db Moving from a farm to another implies just changes in IP address (not name)

Unique dhcp server for all VLANsSupport for DDNS (cr.cnaf.infn.it)

Investigating Quattor for future needs

Our Desired Solution for Resource Access

SHARED RESOURCES among all experiments Priorities and reservations managed by the scheduler

Most of Tier1 computing machines installed as LCG Worker Nodes, with light modifications to support more VOs

Application Software not directly installed on WNs but accessed from outside (NFS, AFS, …)

One or more Resource Manager to manage all the WNs in a centralized way

Standard way to access Storage for each application