CERN Technology Tracking processors memory storage architecures Summary

CERN Technology Tracking processors, memory, storage, architecures Summary of conclusions LCB Workshop, Marseilles, 1 October 1999 Team members: Ingo Augustin, Jean-Philippe Baud, Fabien Collin, Charles Curran, Fabrizio Gagliardi, Frédéric Hemmer, Sverre Jarp, Gordon Lee, Alessandro Marchioro, Bernd Panzer Steindel, Les Robertson, Ben Segal, Rainer Többicke, Pierre Vande Vyvre last update: 21/12/2021 18: 11 les robertson - cern-it-09 -99 1

CERN § processors, memories, basic systems Previous report (1996) was § pessimistic about frequency § optimistic about parallelism (instructions/cycle) § & about right for performance and cost § The Semiconductor Industry Association 1997 roadmap predicts a new generation every 2 years with 0. 1 at 2 GHz appearing around LHC startup § Previous forecasts have been too conservative BUT we start to come close to fundamental limits like thickness of the gate oxide and power dissipation for which novel solutions are required last update 21/12/2021 18: 11 2

CERN Semiconductor Industry Association (SIA) 1997 technology forecast Solution being pursued No known solution last update 21/12/2021 18: 11 3

CERN Processor Cost & Caveats Estimated cost in 2005 § $0. 75 -1. 60 per CERN-Unit CAUTION! § Investments, costs and prices § fabulous fab costs § production volume? § competition? § Achieved parallelism § the best processors only reach 1. 8 ipc on integer codes § will Merced make a significant advance? § compilers may be as important as semiconductors § last update 21/12/2021 18: 11 Home & office PCs are the driving force 4

Secondary Storage CERN § hard disk market is big & healthy § $30 B with 5 -6 large manufacturers s disk d r a h of 75% o PCs t go in areal density should continue to evolve at 60% per year § from 4 Gbits/inch 2 to about 60 Gbits/inch 2 in 2005 § 20 -25 Gbits/inch 2 demonstrated and prices should continue to decrease at 35% per year § $2 -4/GB in 2005 § Technology limits - room for optimism § super-paramagnetic limit currently around 30 -40 Gbits/inch 2 § competitive R&D in materials and heads § alternative technologies using advanced magneto-optics § Data rate per unit of installed capacity - could be a concern § data rate increases only with the linear density § if the mass market does not demand significantly smaller disks we shall have to purchase more capacity than we really need last update 21/12/2021 18: 11 5

CERN Optics & Exotics Paradoxically § now that magnetic recording has caught up with optical recording densities (~5 Gbits/inch 2 with a red laser) § the entertainment business has produced a product which could play an important role in low end computer storage § DVD-RAM - ~9. 4 GB in 2000 § § access time, data transfer rate - no match for hard disk but may well displace magnetic tape in the home & office Holographic storage § development stalled waiting for a new material § active consortium (military interest) mory SS e E M R P ity STOP igh Dens for $50 tra H edit card l U e eel on a cr K e h T B 2. 3 T Nothing else beyond the stage of technology demonstration last update 21/12/2021 18: 11 6

CERN Tertiary Storage Major doubts about the long term § the market driver is backup and archive -write once, read never § the “commodity” market will have strong competition from DVD -- low-end tape devices may disappear BUT - several recent/announced products for the data centre (half inch, linear) 1998 Estimated tape revenue by device class 0. 5"19 mm Helical 6% $5 B market 0. 25" Linear 9% 8 mm Helical 12% 0. 5" Linear 50% 4 mm Helical 23% § STK 9840, DLT 8000, LTO § with strong roadmaps for the next 5 -6 years § 2005 : 20 -50 MB/sec; >100 GB per cartridge Costs § reliable drives unlikely to go below $10 -20 K § robotics - >=$20 per slot (no improvement on today’s best price) last update 21/12/2021 18: 11 7

CERN Tape media costs At LHC media costs will dominate § cannot assume that there will be successors to current product families in the next 5 years § so we may only see factor of 4 improvement in recording density over today’s half inch linear § giving about $0. 50 per GB for raw tape Worst case scenario? § If hard disk is $2/GB and DVD-RAM $0. 20 many applications will abandon tape for the higher functionality of random access storage § Big problem for the magnetic tape business - and HEP We should be very careful about planning for tape as a medium for active data storage last update 21/12/2021 18: 11 8

CERN § § Storage Management Systems Lack of standardisation ==> no exchange of data Limited high end market § missing functionality § product plans / survival unclear § Relatively straightforward HEP requirements § good scalability, predictable performance § Could be a case where a HEP solution would be better and affordable (but surely not multiple HEP solutions!) § Storage Area Networks, Global Shared File Systems § very interesting developments § could mature by 2005, and should be followed § but watch the impact on storage costs last update 21/12/2021 18: 11 9

CERN § Architecture, Interconnects & Clusters Exotic architectures and technologies will have no impact on general purpose computing in the next 5 years § quantum, optical. . multi-threaded (TERA), chaos, . . § games consoles, integrated TV/Internet devices will help to squeeze prices, push performance of regular PCs § § § Expect the distinction between large SMPs, NUMA shared memory and SPP systems to become less clear BUT large integrated systems are likely to remain relatively expensive and be unnecessary for HEP Should see healthy developments in scalable clusters with standardised high performance commodity interconnects § management tools a concern § there are many promising development projects, but they may emphasise high availability/performance rather than high throughput last update 21/12/2021 18: 11 10

full report: http: //nicewww. cern. ch/~les/pasta/welcome. html CERN Summary § § processors OK disks OK - but watch the I/O-rate/capacity ratio § Concern about the future of magnetic tapes it would be wise to consider using disk for active data relegating tape to archive and data exchange § The inherent HEP parallelism makes it realistic to build large computing fabrics from inexpensive components But the management - farms and storage - may require HEP developments Maybe open source may allow this to be done as part of a wider collaboration last update 21/12/2021 18: 11 11