## A heuristic for optimizing stochastic activity networks with applications to statistical digital circuit sizing (2004)

Venue: | IEEE Transactions on Circuits and Systems-I |

Citations: | 12 - 4 self |

### BibTeX

@TECHREPORT{Kim04aheuristic,

author = {Seung-jean Kim and Stephen P. Boyd and Sunghee Yun and Dinesh D. Patil and Mark A. Horowitz},

title = {A heuristic for optimizing stochastic activity networks with applications to statistical digital circuit sizing},

institution = {IEEE Transactions on Circuits and Systems-I},

year = {2004}

}

### OpenURL

### Abstract

A deterministic activity network (DAN) is a collection of activities, each with some duration, along with a set of precedence constraints, which specify that activities begin only when certain others have finished. One critical performance measure for an activity network is its makespan, which is the minimum time required to complete all activities. In a stochastic activity network (SAN), the durations of the activities and the makespan are random variables. The analysis of SANs is quite involved, but can be carried out numerically by Monte Carlo analysis. This paper concerns the optimization of a SAN, i.e., the choice of some design variables that affect the probability distributions of the activity durations. We concentrate on the problem of minimizing a quantile (e.g., 95%) of the makespan, subject to constraints on the variables. This problem has many applications, ranging from project management to digital integrated circuit (IC) sizing (the latter being our motivation). While there are effective methods for optimizing DANs, the SAN optimization problem is much more difficult; the few existing methods cannot handle large-scale problems.

### Citations

3949 |
Convex Optimization
- Boyd, Vandenberghe
- 2004
(Show Context)
Citation Context ...an solve the problem exactly, i.e., globally. If the durations are convex functions of x, then the problem (10) is a convex optimization problem, and so can be (globally) solved very efficiently; see =-=[10, 12, 19, 58, 57]-=-. Another case in which we can solve the problem globally is when the durations are posynomials, or generalized posynomials. In this case the problem (10) is a geometric program (GP), and can be effic... |

2049 |
Numerical Optimization
- Nocedal, Wright
- 1999
(Show Context)
Citation Context ...ns di are. Although we have introduced n + 1 new variables, the constraints are relatively sparse (at least in the variables ti and s), and this can be exploited by a solver for efficiency; see, e.g.,=-=[11, 26, 60, 89]-=-. In many cases we have to settle for a locally optimal solution of the DAN optimization problem (9). But in some cases we can solve the problem exactly, i.e., globally. If the durations are convex fu... |

802 |
Nonlinear programming
- Bertsekas
- 1999
(Show Context)
Citation Context ...ns di are. Although we have introduced n + 1 new variables, the constraints are relatively sparse (at least in the variables ti and s), and this can be exploited by a solver for efficiency; see, e.g.,=-=[11, 26, 60, 89]-=-. In many cases we have to settle for a locally optimal solution of the DAN optimization problem (9). But in some cases we can solve the problem exactly, i.e., globally. If the durations are convex fu... |

782 |
Nonlinear Programming. Athena Scientific, 2nd edition
- Bertsekas
- 1999
(Show Context)
Citation Context ...s di are. Although we have introduced n + 1 new variables, the constraints are relatively sparse (at least in the variables ti and s), and this can be exploited by a solver for efficiency; see, e.g., =-=[11, 28, 62, 91]-=-. In many cases we have to settle for a locally optimal solution of the DAN optimization problem (10). But in some cases we can solve the problem exactly, i.e., globally. If the durations are convex f... |

649 |
Approximation Theorems of Mathematical Statistics
- SERFLING
- 1980
(Show Context)
Citation Context ...For example, if D is Gaussian, then Cmax is the maximum of a number of correlated Gaussian random variables (the path delays), and there are many bounds and asymptotics known for such 6sdistributions =-=[80]-=-. A general and common approach is based on approximating or bounding the distributions of the completion times, by recursively bounding (or approximating) the distribution of completion time of activ... |

299 | Finding the k shortest paths
- Eppstein
- 1999
(Show Context)
Citation Context ... of critical paths, we may wish to leave the set of critical paths described as a subgraph, and not enumerate its paths.) A more sophisticated method can be used to identify the k most critical paths =-=[34]-=-. 2.3 Activity network representation of digital circuits In the remainder of this section we show how activity networks can be used to (approximately) model delay in a digital combinational logic cir... |

282 | Robust Convex Optimization
- Ben-Tal, Nemirovski
- 1998
(Show Context)
Citation Context ...lem to approximately account for some variation in the problem data. A more sophisticated and recent approach to handling uncertainty and variation in optimization problem data is robust optimization =-=[8, 9, 14, 15, 39, 45]-=-. In robust optimization, an explicit model of data uncertainty is used, and the objective is taken to be the average or worst-case value of the objective, over the parameter variation set. Some robus... |

243 | Robust solutions of uncertain linear programs
- Ben-Tal, Nemirovski
- 1999
(Show Context)
Citation Context ...lem to approximately account for some variation in the problem data. A more sophisticated and recent approach to handling uncertainty and variation in optimization problem data is robust optimization =-=[8, 9, 14, 15, 39, 45]-=-. In robust optimization, an explicit model of data uncertainty is used, and the objective is taken to be the average or worst-case value of the objective, over the parameter variation set. Some robus... |

231 | Matching Properties of MOS Transistor
- Pelgrom, Diunmijer, et al.
- 1989
(Show Context)
Citation Context ...x), (16) where α > 0 and γ > 0 are parameters. This simple model scales the relative statistical variation in delay, relative to the mean delay, using a power law. It is inspired by the Pelgrom mode=-=l [68],-=- which predicts a decrease in device parameter variation with increasing device area, due to spatial averaging. The parameter γ gives the relative variation for a minimum sized gate (i.e., xi = 1). W... |

230 |
Stochastic Programming
- Prekopa
- 1995
(Show Context)
Citation Context ...makespan: qα(Cmax) = inf{t : prob(Cmax ≤ t) ≥ α}. In project management, the expected makespan and expected tardiness have been widely used; quantiles are commonly used in finance and risk manag=-=ement [72, 76]-=-. In digital circuit design problems, the probability of missing a deadline (and therefore also the quantiles of Cmax) can be related to yield. In this context, the statistical variation in the durati... |

215 |
Introductory lectures on convex optimization: a basic course
- Nesterov
- 2004
(Show Context)
Citation Context ...an solve the problem exactly, i.e., globally. If the durations are convex functions of x, then the problem (10) is a convex optimization problem, and so can be (globally) solved very efficiently; see =-=[10, 12, 19, 58, 57]-=-. Another case in which we can solve the problem globally is when the durations are posynomials, or generalized posynomials. In this case the problem (10) is a geometric program (GP), and can be effic... |

185 | Statistical Timing Analysis Considering Spatial Correlations Using a Single PERT-like Traversal
- Chang, Sapatnekar
(Show Context)
Citation Context ...riables, possibly dependent. The increasing importance of process variation explains the growing interest in process variation modeling and statistical timing analysis of digital circuits; see, e.g., =-=[4, 16, 21, 42, 44, 47, 63, 64]-=-. While there have been several theoretical and empirical studies of statistical variation in circuit parameters, the statistical modeling of circuit parameter variation is still an active research ar... |

175 |
TILOS: A Posynomial Programming Approach to Transistor Sizing
- Fishburn, Dunlop
- 1985
(Show Context)
Citation Context ...ariables, can be solved. 4.2 Example: cycle time minimization of digital circuits A wide variety of optimization methods have been developed for device and wire sizing in digital circuits; see, e.g., =-=[24, 25, 36, 77, 78, 79]-=-. The cycle time minimization problem can be cast as a DAN optimization problem of the form (9), or an extension that handles rising and falling delays, signal slopes, and so on. As an example, we con... |

162 |
Logical Effort: Designing Fast CMOS Circuits
- Sutherland, Sproull, et al.
- 1999
(Show Context)
Citation Context ...(15) 3 where |Prec(i)| denotes the cardinality of Prec(i), i.e., the number of fan-in gates for gate i. This dependence of ai and bi on the number of fan-in gates is based on the logical effort model =-=[87].-=- The scaling factor 5 in (14) is chosen so that the delay corresponds approximately to the delay, in picoseconds, that would be found in a typical 0.18µm CMOS process. We emphasize again, however, th... |

159 |
Convex analysis and optimization. Athena Scientific
- Bertsekas
(Show Context)
Citation Context ...an solve the problem exactly, i.e., globally. If the durations are convex functions of x, then the problem (11) is a convex optimization problem, and so can be (globally) solved very efficiently; see =-=[10, 12, 21, 60, 59]-=-. Another case in which we can solve the problem globally is when the durations are posynomials, or generalized posynomials. In this case the problem (11) is a geometric program (GP), and can be effic... |

157 |
der Vaart. Asymptotic Statistics
- van
- 1998
(Show Context)
Citation Context ...� max , h k=1 where K : R → R is a probability density function with mean 0 and variance 1, e.g., the standard normal distribution, and h is a positive constant, called the bandwidth of the estimator =-=[90]-=-. Our extensive computational experience with this kernel estimator suggests that a sample of several thousand realizations very often suffices to obtain an accurate estimate of the PDF of Cmax, even ... |

150 | Robust Solutions to Least-Squares Problems with Uncertain Data
- Ghaoui, Lebret
- 1997
(Show Context)
Citation Context ...lem to approximately account for some variation in the problem data. A more sophisticated and recent approach to handling uncertainty and variation in optimization problem data is robust optimization =-=[8, 9, 14, 15, 39, 45]-=-. In robust optimization, an explicit model of data uncertainty is used, and the objective is taken to be the average or worst-case value of the objective, over the parameter variation set. Some robus... |

130 | Robust Discrete Optimization and Network Flows
- Bertsimas, Sim
(Show Context)
Citation Context |

115 |
L-Shaped Linear Programs with Applications to Optimal Control and Stochastic Programming
- Slyke, Wets
- 1969
(Show Context)
Citation Context ...gn for yield (DFY), design for manufacture (DFM), and design centering. Deterministic and stochastic activity network optimization problems arise in many applications, ranging from project management =-=[70, 82, 84]-=- to digital circuit sizing [20]. There are many exact or direct methods for solving DAN optimization problems, even for large-scale problems. There are far fewer methods for solving SAN optimization p... |

112 | Statistical timing analysis for intra-die process variations with spatial correlations
- Agarwal, Blaauw, et al.
(Show Context)
Citation Context ...riables, possibly dependent. The increasing importance of process variation explains the growing interest in process variation modeling and statistical timing analysis of digital circuits; see, e.g., =-=[4, 16, 21, 42, 44, 47, 63, 64]-=-. While there have been several theoretical and empirical studies of statistical variation in circuit parameters, the statistical modeling of circuit parameter variation is still an active research ar... |

98 |
LANCELOT: A Fortran Package for Large-Scale Nonlinear Optimization (Release A
- Toint
- 1992
(Show Context)
Citation Context ...ns di are. Although we have introduced n + 1 new variables, the constraints are relatively sparse (at least in the variables ti and s), and this can be exploited by a solver for efficiency; see, e.g.,=-=[11, 26, 60, 89]-=-. In many cases we have to settle for a locally optimal solution of the DAN optimization problem (9). But in some cases we can solve the problem exactly, i.e., globally. If the durations are convex fu... |

95 | Lectures on modern convex optimization
- Ben-Tal, Nemirovski
- 2000
(Show Context)
Citation Context ...an solve the problem exactly, i.e., globally. If the durations are convex functions of x, then the problem (10) is a convex optimization problem, and so can be (globally) solved very efficiently; see =-=[10, 12, 19, 58, 57]-=-. Another case in which we can solve the problem globally is when the durations are posynomials, or generalized posynomials. In this case the problem (10) is a geometric program (GP), and can be effic... |

95 | An exact solution to the transistor sizing problem ofr CMOS circuits using convex optimization
- Sapatnekar, Rao, et al.
- 1993
(Show Context)
Citation Context ...ariables, can be solved. 4.2 Example: cycle time minimization of digital circuits A wide variety of optimization methods have been developed for device and wire sizing in digital circuits; see, e.g., =-=[24, 25, 36, 77, 78, 79]-=-. The cycle time minimization problem can be cast as a DAN optimization problem of the form (9), or an extension that handles rising and falling delays, signal slopes, and so on. As an example, we con... |

86 | Solving ill-conditioned and singular linear systems: a tutorial on regularization
- Neumaier
- 1998
(Show Context)
Citation Context ...blem; in §3.1.3, the surrogate DAN is used to find a lower bound on a performance measure such as a quantile. This simple heuristic method is similar in spirit to the general method of regularization=-= [59]-=-, in which an extra penalty term is added to a problem to approximately account for some variation in the problem data. A more sophisticated and recent approach to handling uncertainty and variation i... |

84 |
Association of random variables, with applications
- Esary, Walkup
- 1967
(Show Context)
Citation Context ...ntile of the makespan: minimize qα(Cmax) subject to x ∈ Ω. This stochastic programming problem arises in a variety of applications areas, from project management to digital circuit design; see, e=-=.g., [35, 55, 70, 82]-=-. Unfortunately, in all but a few trivial cases, it is very difficult to solve exactly, Numerous methods have been developed for stochastic programming, including random search methods [38], methods b... |

78 | A general probabilistic framework for worst case timing analysis
- Orshansky, Keutzer
(Show Context)
Citation Context ...tion times of its predecessors. Examples can be found in, e.g., [28, 30, 41, 52, 54, 74, 81, 92]. Lower or upper bounds on the expected makespan or expected tardiness in a SAN is considered in, e.g., =-=[13, 17, 64, 91]. -=-In the remainder of this section we describe some relatively simple performance bounds that we will use in the sequel. 3.1.1 The nominal DAN and Jensen’s inequality The DAN (G, ED), with (determinis... |

53 |
Sample-path optimization of convex stochastic performance functions
- Plambeck, S, et al.
- 1996
(Show Context)
Citation Context ...gn for yield (DFY), design for manufacture (DFM), and design centering. Deterministic and stochastic activity network optimization problems arise in many applications, ranging from project management =-=[70, 82, 84]-=- to digital circuit sizing [20]. There are many exact or direct methods for solving DAN optimization problems, even for large-scale problems. There are far fewer methods for solving SAN optimization p... |

52 | Death, taxes and failing chips
- Visweswariah
- 2003
(Show Context)
Citation Context ...2, 7, 53, 62, 85, 94]. As device dimensions shrink, however, growing (relative) statistical uncertainty and process variation will require an approach where design and yield optimization are combined =-=[46, 90, 93].-=- We believe that the method described here is good candidate for such a method. This paper is organized as follows. We describe the deterministic activity network model in §2, and the stochastic acti... |

49 |
Statistical Timing Analysis Using Bounds and Selective Enumeration
- Agarwal, Zolotov, et al.
- 2003
(Show Context)
Citation Context ...as is the makespan. Activity network models have been used in many fields, e.g., project management [27, 32, 33], parallel processing systems [75], and statistical timing analysis of digital circuits =-=[5, 50]-=-. In digital circuit analysis, the durations are usually called delays, and the makespan is called the cycle time or total delay of the circuit. This paper concerns the optimization of a SAN, i.e., th... |

47 | Statistical timing for parametric yield prediction of digital integrated circuits
- Jess, Kalafala, et al.
(Show Context)
Citation Context ...2, 7, 53, 62, 85, 94]. As device dimensions shrink, however, growing (relative) statistical uncertainty and process variation will require an approach where design and yield optimization are combined =-=[46, 90, 93].-=- We believe that the method described here is good candidate for such a method. This paper is organized as follows. We describe the deterministic activity network model in §2, and the stochastic acti... |

44 | Impact of spatial intrachip gate length variability on the performance of high-speed digital circuits
- Orshansky, Milor, et al.
(Show Context)
Citation Context ...ur mean delay model, this statistical model is chosen for simplicity, not accuracy. In any case, modeling the statistics of gate delay is an area of active research, with many open issues; see, e.g., =-=[51, 61, 65, 66]. For th-=-e optimization problems, we imposed a constraint on the sum of the gate multipliers, as well as lower and upper bounds: n� xi ≤ 1000, 1 ≤ xi ≤ 100, i = 1, . . . , n. (18) i=1 Assuming the gate... |

39 | A family of adders
- Knowles
- 2001
(Show Context)
Citation Context ...earch area; so far, no consensus has emerged as to what the best models are. 3.3.1 Example: Ladner-Fisher 32-bit adder As an example, we carry out MC timing analysis of the Ladner-Fisher 32-bit adder =-=[49]-=-, which is used throughout the paper to illustrate our main points. This circuit consists of 461 gates (n = 461), 64 input gates, 32 output gates, and 1714 arcs. The associated DAG G = (N, A) has a to... |

34 |
Convex majorization with an application to the length of critical paths
- Nadas
(Show Context)
Citation Context ...hrough the activity; it is the sum of the criticality indexes of the paths containing the activity. Identifying activities and paths with large criticality index is a well studied problem (see, e.g., =-=[18, 23, 29, 54]). Sev-=-eral scalar performance measures can be used to characterize the makespan Cmax of a SAN. Examples include: • Expected makespan: E(Cmax). • Probability of missing a deadline T0: prob(Cmax ≥ T0). ... |

32 |
Statistical Timing Analysis of Combinational Logic Circuits
- Jyu, Malik, et al.
- 1993
(Show Context)
Citation Context ...riables, possibly dependent. The increasing importance of process variation explains the growing interest in process variation modeling and statistical timing analysis of digital circuits; see, e.g., =-=[4, 16, 21, 42, 44, 47, 63, 64]-=-. While there have been several theoretical and empirical studies of statistical variation in circuit parameters, the statistical modeling of circuit parameter variation is still an active research ar... |

31 | Digital circuit optimization via geometric programming
- Boyd, Kim, et al.
- 2005
(Show Context)
Citation Context ... 38, 79, 80, 81]. The cycle time minimization problem can be cast as a DAN optimization problem of the form (10), or an extension that handles rising and falling delays, signal slopes, and so on. See =-=[19]-=- for more on GP-based sizing of digital circuits. As an example, we consider the 32-bit Ladner-Fisher adder circuit. Each of the 459 gates can be scaled by a multiplier or scale factor xi ≥ 1, which a... |

28 |
Bounding the Project Completion Time Distributions in PERT Networks
- Dodin
- 1985
(Show Context)
Citation Context ...or approximating) the distribution of completion time of activity i, given bounds (or approximations) of the distributions of the completion times of its predecessors. Examples can be found in, e.g., =-=[28, 30, 41, 52, 54, 74, 81, 92]-=-. Lower or upper bounds on the expected makespan or expected tardiness in a SAN is considered in, e.g., [13, 17, 64, 91]. In the remainder of this section we describe some relatively simple performanc... |

28 |
Computer-aided design for VLSI circuit manufacturability
- MALY
- 1990
(Show Context)
Citation Context ...ocess variation can be indirectly handled by incorporating simple margins 1sin the timing and other critical constraints, or by a post-design step that does centering or yield improvement; see, e.g., =-=[1, 2, 7, 53, 62, 85, 94]-=-. As device dimensions shrink, however, growing (relative) statistical uncertainty and process variation will require an approach where design and yield optimization are combined [46, 90, 93]. We beli... |

27 | Gradient-based optimization of custom circuits using a static-timing formulation
- Conn, Elfadel, et al.
- 1999
(Show Context)
Citation Context ...ariables, can be solved. 4.2 Example: cycle time minimization of digital circuits A wide variety of optimization methods have been developed for device and wire sizing in digital circuits; see, e.g., =-=[24, 25, 36, 77, 78, 79]-=-. The cycle time minimization problem can be cast as a DAN optimization problem of the form (9), or an extension that handles rising and falling delays, signal slopes, and so on. As an example, we con... |

26 | 2003): A Robust Optimization Approach to Supply Chain Management. Working Paper, MIT
- Bertsimas, Thiele
(Show Context)
Citation Context |

22 |
LOQO user’s manual
- Vanderbei
- 1992
(Show Context)
Citation Context |

21 | Characterization of spatial intrafield gate cd variability, its impact on circuit performance, and spatial mask-level correction
- Orshansky, Milor, et al.
- 2004
(Show Context)
Citation Context ...ur mean delay model, this statistical model is chosen for simplicity, not accuracy. In any case, modeling the statistics of gate delay is an area of active research, with many open issues; see, e.g., =-=[51, 61, 65, 66]. For th-=-e optimization problems, we imposed a constraint on the sum of the gate multipliers, as well as lower and upper bounds: n� xi ≤ 1000, 1 ≤ xi ≤ 100, i = 1, . . . , n. (18) i=1 Assuming the gate... |

20 | Uncertainty-aware circuit optimization
- Bai, Visweswariah, et al.
(Show Context)
Citation Context ...ocess variation can be indirectly handled by incorporating simple margins 1sin the timing and other critical constraints, or by a post-design step that does centering or yield improvement; see, e.g., =-=[1, 2, 7, 53, 62, 85, 94]-=-. As device dimensions shrink, however, growing (relative) statistical uncertainty and process variation will require an approach where design and yield optimization are combined [46, 90, 93]. We beli... |

20 |
Tau: Timing analysis under uncertainty
- Bhardwaj, Vrudhula, et al.
- 2003
(Show Context)
Citation Context |

19 |
Some Analysis Techniques for Asynchronous Multiprocessor Algorithms
- Robinson
(Show Context)
Citation Context ...N), the durations of the activities are random variables, as is the makespan. Activity network models have been used in many fields, e.g., project management [27, 32, 33], parallel processing systems =-=[75]-=-, and statistical timing analysis of digital circuits [5, 50]. In digital circuit analysis, the durations are usually called delays, and the makespan is called the cycle time or total delay of the cir... |

16 | A Computational Study on Bounding the Makespan Distribution in Stochastic Project Networks
- Ludwig, Möhring, et al.
(Show Context)
Citation Context ...or approximating) the distribution of completion time of activity i, given bounds (or approximations) of the distributions of the completion times of its predecessors. Examples can be found in, e.g., =-=[28, 30, 41, 52, 54, 74, 81, 92]-=-. Lower or upper bounds on the expected makespan or expected tardiness in a SAN is considered in, e.g., [13, 17, 64, 91]. In the remainder of this section we describe some relatively simple performanc... |

15 |
Conditional Monte Carlo: A simulation technique for stochastic network analysis
- Burt, Garman
- 1971
(Show Context)
Citation Context ... we describe basic Monte Carlo analysis of a SAN; more sophisticated methods can be used to get higher accuracy with fewer samples, or combine Monte Carlo simulation with bounding methods (see, e.g., =-=[22, 73, 83, 86]-=-). More discussion of sampling strategies for Monte Carlo analysis of SANs can be found in [6, 28, 41, 54, 74, 92]. In basic Monte Carlo analysis, we draw M independent samples d (1) , . . . , d (M) f... |

15 |
Jiytune: circuit optimization using time-domain sensitivities
- Conn, Coulman, et al.
- 1997
(Show Context)
Citation Context |

15 | Analysis for Delay Defects and Noise Effects in Deep Submicron Devices
- Liou
- 2003
(Show Context)
Citation Context ...ur mean delay model, this statistical model is chosen for simplicity, not accuracy. In any case, modeling the statistics of gate delay is an area of active research, with many open issues; see, e.g., =-=[51, 61, 65, 66]. For th-=-e optimization problems, we imposed a constraint on the sum of the gate multipliers, as well as lower and upper bounds: n� xi ≤ 1000, 1 ≤ xi ≤ 100, i = 1, . . . , n. (18) i=1 Assuming the gate... |

15 |
On uncertainty, ambiguity, and complexity in project management
- Pich, Loch, et al.
- 2002
(Show Context)
Citation Context ...e all random variables. We will denote the makespan (random variable) as Cmax. Stochastic activity networks were originally developed in the late 1950s for analyzing project management and scheduling =-=[69]-=-. Most work on project management has focused on project scheduling techniques such as project evaluation and review techniques (PERT) and the critical path method (CPM) (see, e.g., [27, 32, 33]). We ... |

14 |
The completion times of PERT networks
- Robillard, Trahan
- 1976
(Show Context)
Citation Context ...or approximating) the distribution of completion time of activity i, given bounds (or approximations) of the distributions of the completion times of its predecessors. Examples can be found in, e.g., =-=[28, 30, 41, 52, 54, 74, 81, 92]-=-. Lower or upper bounds on the expected makespan or expected tardiness in a SAN is considered in, e.g., [13, 17, 64, 91]. In the remainder of this section we describe some relatively simple performanc... |