Mallach, Sven ORCID: 0000-0001-5335-0678 (2015). Exact Integer Programming Approaches to Sequential Instruction Scheduling and Offset Assignment. PhD thesis, Universität zu Köln.
|
PDF
diss_mallach_kups.pdf - Published Version Download (2MB) |
Abstract
The dissertation at hand presents the main concepts and results derived when studying the optimal solution of two NP-hard compiler optimization problems, namely instruction scheduling and offset assignment, by means of integer programming. It is the outcome of several years of research as an assistant at Michael Jünger's computer science chair in Cologne, with the particular aim to apply exact mathematical optimization techniques to real-world problems arising in the domain of technical computer science. The two problems studied are rather unrelated apart from the fact that they both take place during the machine code generation phase of a compiler and deal with the handling of limited resources. Instruction scheduling is about the assignment of issue clock cycles to instructions in the presence of precedence, latency, and resource constraints such that the total time needed to execute all the instructions is minimized. Offset assignment deals with storage layouts of program variables and the efficient use of address registers for accesses to these variables. The objective is to employ specialized instructions in order to minimize the overhead caused by address computations. While instruction scheduling needs to be carried out by almost every present compiler irrespective of the processor architecture, the offset assignment problem occurs mainly in compilers for highly specialized processor designs. Instruction scheduling is a well-studied field where several exact and heuristic approaches have been developed and experimentally evaluated in the past. In this thesis, we concentrate on the basic-block instruction scheduling problem for single-issue processors. Basic blocks are program fragments with no side-entrances and -exits, i.e., every instruction of a basic block needs to be executed before the control flow may leave it and enter another basic block. Single-issue processors are capable of starting the execution of exactly one instruction per clock cycle. A number of techniques to preprocess instances of the basic-block instruction scheduling problem were proposed in the literature and are, with emphasis on the more recent ones that arose since the year 2000, thoroughly reviewed in this thesis. They finally led to a constraint programming approach in 2006 that was shown to solve about 350,000 instances to optimality and where some of these instances comprised up to about 2,500 instructions. The last attempt to tackle the problem using integer programming however dates to a time prior to the publication of the latest preprocessing advances. While being successful on a set of instances that impose very restrictive latency constraints, it was shown to be unable to solve hundreds of instances from the aforementioned benchmark set that comprises also large and varying latencies. In addition, the previous integer programming models were almost all based on so-called time-indexed formulations where decision variables model an explicit assignment of instructions to clock cycles. In this thesis, a completely different and novel approach is taken based on the linear ordering problem, a well-studied combinatorial optimization problem. The new models lead to alternative characterizations of the feasible solutions to the basic-block instruction scheduling problem. These facilitate the employment of advanced integer programming methodologies, in particular the design of branch-and-cut algorithms that can handle larger instances. The formulations are further extended by additional inequalities that can be used as cutting planes. Combined with the preprocessing routines that are partially extended and improved as well, the respective solver implementation eventually turned out to be competitive to the constraint programming method. Reaching this point has taken some years and this thesis presents not only the derived models but also several ideas and byproducts that arose in the meantime, and that can help and inspire researchers even if they aim at the application of different solution methodologies. The starting point regarding the offset assignment problem was a different one because especially exact solution approaches were rather rare prior to the models presented in this thesis. The offset assignment problem arose in the 1990s and is considered in several variants that are of theoretical and practical interest. In the simplest one, a processor is assumed to provide only a single address register and only very restricted possibilities to avoid address computation overhead. However, even this simplest variant, that may serve as a building block for the more complex ones, is already NP-hard and has been studied mainly from a heuristic point of view. The few existing exact solution approaches were not capable to solve moderately sized instances so that the quality of heuristic solutions relative to the optimum was hardly known at all. Again, the inspection of the combinatorial structure of the various problem variants turned out to be the key for designing branch-and-cut implementations that can profit from knowledge about related combinatorial optimization problems. The implementation targeting the simple problem variant was the first capable to optimally solve the majority of about 3,000 instances collected in a standard benchmark set. The method could then be further generalized in two steps. First, in a collaboration with Roberto Castañeda Lozano, additional techniques could be incorporated into the approach in order to handle multiple address registers. Fortunately, the methods could then even be further extended to as well deal with more flexible addressing capabilities. In this way, the thesis at hand does not only answer the question how large the address computation overhead can be when using heuristics, but as well presents first results that allow to analyze the impact of the mentioned increased addressing capabilities on the runtime performance and size of real-world programs.
Item Type: | Thesis (PhD thesis) | ||||||||
Translated abstract: |
|
||||||||
Creators: |
|
||||||||
URN: | urn:nbn:de:hbz:38-60399 | ||||||||
Date: | 16 March 2015 | ||||||||
Language: | English | ||||||||
Faculty: | Faculty of Mathematics and Natural Sciences | ||||||||
Divisions: | Faculty of Mathematics and Natural Sciences > Department of Mathematics and Computer Science > Institute of Computer Science | ||||||||
Subjects: | Natural sciences and mathematics Mathematics |
||||||||
Uncontrolled Keywords: |
|
||||||||
Date of oral exam: | 15 January 2015 | ||||||||
Referee: |
|
||||||||
Refereed: | Yes | ||||||||
URI: | http://kups.ub.uni-koeln.de/id/eprint/6039 |
Downloads
Downloads per month over past year
Export
Actions (login required)
View Item |