Portal de Periódicos da CAPES

Computational intermediation and the evolution of computation as a commodity

2004; Taylor & Francis; Volume: 36; Issue: 11 Linguagem: Inglês

10.1080/0003684042000247334

1466-4283

Antony Davies,

Cellular Automata and Applications

Abstract The consumer who purchases computational power ultimately purchases a reduction in the time interval between the initiation and the completion of work. This paper looks at computation as a commodity and the nascent industry of computational intermediation, and proposes a model for the market for computational power as distinct from the market for computers. Some interesting results emerge. The model implies that the demand for computation is discontinuous and that there is a lower limit to the quantity of computation consumers will demand that is independent of the price of power. The model identifies a range of computational powers that could be supplied by computational intermediaries but which will not be supplied by computer manufacturers, and suggests a model for pricing computation. Acknowledgements I thank Steven L. Armentrout, James Gannon, and Renate Neely Brown for comments and criticisms on an earlier version of this work. All remaining errors and omissions are my own. Notes Among the entrepreneurs was a young Ross Perot who, with the founding of Electronic Data Systems, was one of the first to offer computation for rent. Interactive Week, 15 January 2001, p. 74. Clock-speed (i.e. cycles per second, or Hz) is frequently used as a measure for processor power in PCs. The measure is an imperfect power metric because the number of cycles required to perform a given mathematical operation varies among chip designs (this is why, for example, an 800 MHz Pentium chip is faster than an 800 MHz Celeron chip). A preferred power metric is flops (floating point operations per second). This metric directly measures the number of mathematical operations the chip performs per unit time. While a preferable metric to clock-speed, flops is also imperfect because (a) it is representative of effective speed only for operations that involve floating point calculations, and (b) the results obtained can vary depending on the type of mathematical tests performed. Throughout this paper, unless otherwise noted, the maximum sustained flops ratings are used (vs. peak flops rating) obtained from a Linpack inversion of a 1000 × 1000 matrix. Cf. Dongarra (Citation2002). A Dell Pentium 3, 550 MHz operates at a sustained maximum power of 0.08 GF. The peak operating power – which is often quoted but rarely achieved – is 0.55 GF. Cf. The Performance Database Server, www.netlib.org. Cf. www.top500.org, June 2002. Within the industry, the legacy terms ‘super computation (computer)’ have been supplanted by ‘high-performance computation (computer),’ or ‘HPC.’ Roberts (Citation2000). Baskett and Hennessey (Citation1993). Cf. Hawick et al. (Citation1998). International Data Corporation, Report #W20058, September 1999. Chris Willard, Vice President of Research for IDC, indicates that the sales numbers in W20058 understates true sales for 1993 and 1994. Davies (Citation2001b). The legacy term for ‘cluster computer’ is ‘Beowulf cluster,’ taken from the name of one of the first cluster computers. See www.beowulf.org for a synopsis of the early history of cluster computers. Wasson (Citation2001).www.tech-report.com As of January 2002, a headless, 1.8 GHz, Dell Pentium 4 with 256 SDRAM retailed for $928. The Linpack rating (matrix size = 1000) for this processor is approximately 0.24 GF. Examples include (numbers are shown in present value, annual cost terms, adjusting for inflation and Moore's Law): The Danforth Cluster (www.danforthcenter.org) at $5,700 per GF, Loki (loki-www.lanl.gov) at $4,500 per GF, SWARM (www.cs.orst.edu) at $5,100 per GF, and Brahma (www.phy.duke.edu) at $4,100 per GF. Joseph et al. (Citation2000c). Price–performance numbers are usually quoted in terms of purchase price. The actual annual cost of a cluster computer (including amortized purchase price, infrastructure, maintenance, and support) is around 125% of the purchase price of the nodes ( Davies, Citation2001a). See www.setiathome.ssl.berkeley.edu. See www.top500.org. Note that several power figures are listed for the Top 500 computers. Among them are the maximum sustained power and the peak power. In most public relations literature, peak power is quoted. Peak power has been described as ‘the lowest speed the machine is guaranteed not to achieve.’ Throughout this paper, the maximum sustained power is used. Note that, on average, peak power is 1.5 times sustained maximum power. ASCI White's purchase-price–performance ratio is on a par with that of traditional HPC because ASCI White is, strictly speaking, a hybrid machine. It is a cluster computer whose nodes are not off-the-shelf PCs, but are themselves low-power traditional HPC. Thus, ASCI White's price–performance ratio is on a par with traditional HPC. What ASCI White gains through clustering is a power level unattainable by traditional HPC. Cf. http://www-1.ibm.com/servers/eserver/pseries/hardware/largescale/ supercomputers/ asciwhite/ See IBM (www-1.ibm.com/servers/events/grid.html), EGrid (www.egrid.org), Intel (www.intel.com/ids/p2p), Sun (www.sun.com/software/gridware), Parabon Computation (www.parabon.com), DataSynapse (www.datasynapse.com), United Devices (www. uniteddevices.com), and Entropia (www.entropia.com). See, for example, www.computeagainstcancer.org, and www.fightaidsathome.org. Sun and Datasynapse offer SDK's for ‘enterprise computing grids.’ An enterprise computing grid is like an Internet computing grid with the exception that all of the computers on the grid are ‘in-house.’ Thus, an enterprise computing grid is a middle-step between a cluster computer and an Internet computing grid: the consumer of the computation owns the computers, but the computation generated is idle capacity. In the case of an enterprise computing grid, the consumer does incur infrastructure and amortization costs but, because the grid harnesses idle capacity (and ignoring the cost of the software that administrates the grid), the marginal increase in infrastructure and amortization from establishing the grid is zero. For example, from a laptop computer connected to the Internet via a phone line, a statistical program has been run on Parabon Computation's platform that was distributed to 5000 PCs. Thirty-six hours later, the entire job had completed and the results were downloaded from Parabon's server. Run on a single computer, the job would have required ten years to complete. Because GF are measured on a per-unit-time basis (i.e. floating-point operations per second), multiplying by years cancels the time dimension. Thus, 1 GFY is approximately 3.15 × 1016floating-point operations. Gordon Moore, co-founder of Intel, predicted in 1965 that the number of transistors per square inch (called data density) on integrated circuits would double every year. The pace of increase in data density subsequently slowed to a doubling every 18 months. Note that Moore's Law implies δ δ would continue to hold. Nordhaus (Citation2002). Brown (Citation2000), Ezzell (Citation2000), Howard (Citation2000), Langreth and Adams (Citation2000), Meieran (Citation2000) and Thayer (Citation2000). Joseph et al. (Citation2000d) and Pescovitz (2000). The solution in EquationEquation 25 reduces to the solution in EquationEquation 8 for b = 2, and . In the limit, as b → 1, EquationEquation 28 approaches Ω>aW. Joseph et al. (Citation2000a). Nordhaus (Citation2002). Note that inflation does not matter because it impacts both k and Ω. This assumes that the nodes are stacked in a rack configuration. This configuration occupies the minimum possible floor space. Spector (Citation2000) and ‘Case in point: power plant’, Genome Technology, May 2001. This is based on an upfront cost of $100 to $150 per square foot amortized at a 15% cost of capital over 15 years. Genome Technology, May 2001, p. 22. In the steady-state, a 450 MHz Pentium III draws 0.3 amps. This power consumption excludes the monitor because, in a cluster computer, the nodes do not require monitors. For example, Genetic Programming, Inc. (www.genetic-programming.com) has a 1000 node (450 MHz) cluster that requires two 24-ton air conditioners. At 7 cents per kWh the cost of air conditioning is $123 per ton annually (www.eren.doe.gov). For example, software for the 1,040 processor Danforth Cluster costs $800 000 per year (Genome Technology, May 2001, p. 22). Assuming one full-time IT professional per 300 nodes at an annual salary of $70 000 plus 20% employer burden. Spector (Citation2000); ‘Case in point: power plant,’ Genome Technology May 2001; ‘High-Performance Computing Industry Changes,’ International Data Corporation, Telebriefing, 11 May 2001. As of January 2002, a headless, 1.8 GHz Dell Pentium 4 retailed for $928 and generates 0.24 GF. This is 125% of the reported purchase price of the nodes. In the case of ASCI White, the reported cost of $110 million is taken as including installation, infrastructure, support, and maintenance. Present value calculations assume a useful life of three years. This assumes that the machines are operated 85% of the time. The partial utilization reflects down time for periodic maintenance and upgrades. Hawick et al. (Citation1998). See www.llnl.gov/asci/news/white_news.html. See www.danforthcenter.org for details on the cluster. These firms have asked not to be identified. The financial institution is a Global 200 corporation. Genomics Firm #1 is one of the leading genomics companies in the US. Genomics Firm #2 is a medium-sized firm with sales in 2000 of $150 million. This company's cluster computer is comprised of Alpha computers. Alphas have a significantly greater price–performance ratio.