High-performance I/O and its implication to computer architecture

Is I/O inherently inefficient or is it how we use it? Historically, I/O has been the bottleneck of computing systems. While costs of rotating storage have come down, disks are largely governed by physical laws where mechanical constraints tend to throttle performance compared to other system elements. By taking a fresh look at the role of disks and in particular, how we map and store data, we can achieve significant gains in performance.; When faced with an I/O problem, the conventional approach is to simply add more spindles. However, we eventually reach a limit: the ability of an I/O bus to transport the data to the requesting host(s) or the host processor's inability to store and process the data. Hence, the I/O wall. To address this problem researchers have proposed systems that process data in place by using intelligent disks (I-Disks).; In this work we examine the overall efficiency of the I/O operation. To that end, we define a figure of merit, FOM, that is defined to be the efficiency of the data access, $x$, times the utility of that access, $y$. FOM=x^*y . ; While we never achieve an optimum value of unity, It is a useful metric and convenient for measuring relative performance. Traditional ad-hoc decision-support type of applications will have FOM values on the order of .001–.005 due to the randomness inherent in the access and the fact that often we are often reading more data than we need.; Given disk parameters (latency and seek time) are fixed and inescapable, we change the order of processing so that what normally would be a limiting factor, becomes largely transparent. Operations are atomic to the cylinder and synchronous. Thus, the FOM approaches unity.; Drawing from several benchmarks, we compare relative performance and show that FOM tracks the relative performance rather closely and is a reasonable predictor of performance. We validated our analytical model using a prototype system.; In summary, by organizing data so as to preserve the logical topology and striving for coherent data access significant performance gains can be achieved in applications heretofore I/O constrained.