OPTIMIZING THE JAVA VIRTUAL MACHINE INSTRUCTION SET BY ...

More documents

Recommendations

Info

$LATEX for Philosophers - University of Calgary$

40 3.3 Profiling Much of the work presented in this thesis makes use of data collected by profiling Java applications. Section 3.3.1 describes some mechanisms that are available for collecting profile data for Java applications. Other studies that collected and evaluated profile data from Java applications are discussed in Section 3.3.2. 3.3.1 Gathering the Profile Data All of the profile data used during the studies presented in this thesis was collected by modifying an existing Java Virtual Machine. These modifications introduced the data structures necessary to store the collected profile data and the hooks necessary to detect events of interest when they occurred. The choice to profile the applications in this way was made because implementing multicodes already necessitated making modifications to the Java Virtual Machine. As a result, the structure of the Java Virtual Machine and the location of the required data was already known. Collecting the profile data in this manner also offered maximum flexibility, providing raw, unfiltered access to the internal workings of the virtual machine as the application ran. However, other options for profiling Java applications do exist. These options are briefly discussed in the following subsections. The JVM Tool Interface The Java Virtual Machine Tool Interface (JVMTI) [106] is a new addition to version 1.5.0 of the Java platform. It replaces the Java Virtual Machine Profiling Interface (JVMPI), removing many of the limitations imposed by the older interface [86]. This interface is intended for the development of tools that monitor or record the state of the Java Virtual Machine as it executes applications. It also provides the functionality necessary to develop debugging, thread analysis and coverage analysis tools. The JVMTI provides a two-way interface. An application that uses the JVMTI to examine the virtual machine, referred to as an agent, can query the status of the virtual machine directly using functions. An agent can also be notified when certain changes within the virtual machine occur through events. Agent applications can be developed in C, C++ or any language which supports C calling conventions. This interface is related to the material presented in this thesis in that it provides the capability to collect a large amount of information about the bytecodes executed by the application. In particular, it can be employed in such a manner so as to record
41 each bytecode executed as was accomplished in Chapter 5 and Chapter 7. In addition, the interface can be used to record the operand values provided to those bytecodes as was accomplished in Chapter 6. However, this interface does not provide the capability to modify the instruction set for the virtual machine which is functionality critical to the implementation of multicodes, discussed in Chapter 7. The JVM Profiler Interface In conjunction with the JVM Debug Interface [105], the JVM Profiler Interface [104] provides functionality that is similar to that of the newer JVM Tool Interface. The JVMPI was introduced in version 1.1 of the Java Software Development Kit as an experimental profiling interface. It retained that status when it was ported to the HotSpot virtual machine in version 1.3.0. Like the JVMTI, the JVMPI provides sufficient functionality to perform the profiling activities described in Chapters 5, 6 and 7. However, the JVMPI does not provide functionality for adding new bytecodes. Consequently, it can only be used to aid in the profiling required in order to identify of multicodes. It is not possible to use the JVMPI to implement multicodes without making other modifications to the virtual machine. Profiling Tools The Sable Research Group at McGill University has developed a profiling tool known as *J which gathers information about Java applications as they execute [46]. This tool, which makes use of the JVMPI to collect its data, includes a trace generator which records the application’s behaviour. It also consists of a trace analyzer which analyzes the trace data to generate dynamic metrics and store them in a database. Examples of the metrics that can be reported include the number of bytecodes touched during the execution of an application and the average number of distinct receiver objects per call site among many others. It isn’t clear if *J is capable of recording statistics related to how many times each opcode is executed or the information necessary to identify the sequences of bytecodes executed in its as-distributed form. However, it was designed with extensibility in mind, suggesting that adding this functionality would not be overly difficult if it is not already supported. Many additional high level profiling tools are available. These tools concentrate on collecting profile data that will help identify performance bottlenecks within the
Page 1 and 2:
OPTIMIZING THE JAVA VIRTUAL MACHINE
Page 3 and 4:
Abstract Since its public introduct
Page 5 and 6:
Acknowledgments While my name is th
Page 7 and 8: 2.4.4 Direct Threading . . . . . .
Page 9 and 10: 7.2 An Overview of Multicodes . . .
Page 11 and 12: List of Figures 2.1 Core Virtual Ma
Page 13 and 14: 7.11 201 compress Performance by Mu
Page 15 and 16: B.1 The Steps Followed to Evaluate
Page 17 and 18: 6.1 Most Frequently Occurring Const
Page 19 and 20: 1 Chapter 1 Introduction When Moore
Page 21 and 22: 3 Studies were conducted that consi
Page 23 and 24: 5 Chapter 2 An Overview of Java Thi
Page 25 and 26: 7 Addressing this need led to the d
Page 27 and 28: 9 performs garbage collection, a cl
Page 29 and 30: 11 also has the ability to contain
Page 31 and 32: 13 ... i − 1 iload 0x04 dstore 0x
Page 33 and 34: 15 ... aload 3 iconst 2 faload ...
Page 35 and 36: 17 Condition Equal Unequal Greater
Page 37 and 38: 19 transfered to the start of a new
Page 39 and 40: 21 piled into machine language inst
Page 41 and 42: 23 Figure 2.6: Interpreter Engine C
Page 43 and 44: 25 Figure 2.8: Interpreter Engine C
Page 45 and 46: 27 The following section examines t
Page 47 and 48: 29 execute (JMethod * method, Slot
Page 49 and 50: 31 codelets shown in this example a
Page 51 and 52: 33 3.1 Instruction Set Design The a
Page 53 and 54: 35 is replaced with a new opcode if
Page 55 and 56: 37 method. However, the Java Virtua
Page 57: 39 3.2.5 Bytecode Optimization Tool
Page 61 and 62: 43 executing the application. The a
Page 63 and 64: 45 Chapter 4 Despecialization The J
Page 65 and 66: 47 ... i − 1 iload 3 iload 0x04 i
Page 67 and 68: 49 purpose bytecodes. The notation,
Page 69 and 70: 51 is present in the constant pool
Page 71 and 72: 53 ... iload 3 iflt ... i − 1 i i
Page 73 and 74: 55 the stack challenging in some ca
Page 75 and 76: Figure 4.7: Despecialization of Bra
Page 77 and 78: 59 Specialized Bytecode load store
Page 79 and 80: 61 tual machine (RVM). Version 2.3.
Page 81 and 82: 63 Test Total Weighted Condition Pe
Page 87 and 88: 69 4.2.6 Performance Summary Compar
Page 89 and 90: 71 Category Original Desp. Percent
Page 91 and 92: 73 the maximum value encountered af
Page 93 and 94: 75 • Efficient support of new lan
Page 95 and 96: 77 difference in the performance of
Page 97 and 98: 79 JGF RayTracer: A 150 by 150 pixe
Page 99 and 100: 81 # Average % of Specialized Despe
Page 101 and 102: 83 provided by only two vendors. IB
Page 103 and 104: Figure 5.1: Average Performance Res
Page 105 and 106: 87 Figure 5.4: Performance Results
Page 107 and 108: 89 Figure 5.8: Performance Results
Page 109 and 110:
91 ences in background processes, m
Page 111 and 112:
93 difference in performance due to
Page 113 and 114:
95 when a pure interpreter virtual
Page 115 and 116:
97 It is observed that many of the
Page 117 and 118:
99 Within this category, it was als
Page 119 and 120:
101 machine code for the idioms. As
Page 121 and 122:
103 were performed was less than 0.
Page 123 and 124:
105 these despecializations in a me
Page 125 and 126:
107 Chapter 6 Operand Profiling Cha
Page 127 and 128:
109 the applications. In addition,
Page 129 and 130:
111 The aload 0 bytecode was the mo
Page 131 and 132:
113 function being integrated. Beca
Page 133 and 134:
115 Implementing alignment rules un
Page 135 and 136:
117 these data types in the list of
Page 137 and 138:
Figure 6.3: Distribution of Local V
Page 139 and 140:
121 are never executed by any of th
Page 141 and 142:
123 load and store bytecodes. It is
Page 143 and 144:
125 the time because the index bein
Page 145 and 146:
127 6.2.5 Fields Only four bytecode
Page 147 and 148:
129 Figure 6.8: Data types Accessed
Page 149 and 150:
131 lar, the new set of specialized
Page 151 and 152:
133 multicode block is defined to b
Page 153 and 154:
135 Figure 7.1: Two Bytecodes and t
Page 155 and 156:
137 because of the presence of the
Page 157 and 158:
139 cache performance. As a result,
Page 159 and 160:
141 Block 1: Block 2: aload 0 → g
Page 161 and 162:
143 Sequence Count Score getfield a
Page 163 and 164:
Figure 7.3: Iterative Multicode Ide
Page 165 and 166:
147 number of bytecodes that will b
Page 167 and 168:
149 BlockList: A list of the byteco
Page 169 and 170:
151 Both the Total Bytecodes Score
Page 171 and 172:
153 ... i − 1 ❄ ifeq 0x00 0x15
Page 173 and 174:
155 Figure 7.7: Number of Unique Ca
Page 175 and 176:
Figure 7.9: Cumulative Multicode Bl
Page 177 and 178:
159 However, while considering long
Page 179 and 180:
161 Figure 7.10: 201 compress Perfo
Page 181 and 182:
163 Figure 7.14: Length 209 db Perf
Page 183 and 184:
165 Figure 7.18: 222 mpegaudio Perf
Page 185 and 186:
167 7.6.2 Overall Performance Acros
Page 187 and 188:
169 a larger portion of the instruc
Page 189 and 190:
171 // iload_1: stack[sp+1] = local
Page 191 and 192:
173 // iload_1: // iconst_1: // iad
Page 193 and 194:
175 Such bytecodes have a variable
Page 195 and 196:
177 ∆ x1 /x 2 /x 3 /.../x n→ an
Page 197 and 198:
179 Popping any values left on the
Page 199 and 200:
181 7.9 Conclusion This chapter has
Page 201 and 202:
183 ated will differ from that occu
Page 203 and 204:
185 Figure 8.1: A Comparison of the
Page 205 and 206:
187 performed impacts the sequences
Page 207 and 208:
189 Figure 8.2: 201 compress Perfor
Page 209 and 210:
191 Figure 8.5: 213 javac Performan
Page 211 and 212:
193 Compared to the complete despec
Page 213 and 214:
195 Figure 8.10: 209 db Performance
Page 215 and 216:
197 performing multicode substituti
Page 217 and 218:
199 Chapter 9 An Efficient Multicod
Page 219 and 220:
201 alphabet employed by each of th
Page 221 and 222:
9.4 An Algorithm for the Most Contr
Page 223 and 224:
Figure 9.3: The Impact of Algorithm
Page 225 and 226:
207 Algorithm NonOverlappingScore(n
Page 227 and 228:
209 Algorithm CountUniqueOccurrence
Page 229 and 230:
211 this algorithm did not consider
Page 231 and 232:
213 exceptions if the value being c
Page 233 and 234:
215 Figure 10.1: Multicode Blocks u
Page 235 and 236:
217 approximately 30 percent of the
Page 237 and 238:
219 a strategy other than direct th
Page 239 and 240:
221 10.2.6 Combining Multicodes and
Page 241 and 242:
223 10.2.9 Expanding the Class File
Page 243 and 244:
225 revealed that none of the const
Page 245 and 246:
227 based on these profile results
Page 247 and 248:
229 stitution were performed. It co
Page 249 and 250:
231 [13] B. Alpern, C. R. Attanasio
Page 251 and 252:
233 [33] K. Casey, D. Gregg, M. A.
Page 253 and 254:
235 [55] D. Gregg and J. Waldron. P
Page 255 and 256:
237 [78] E. M. McCreight. A space-e
Page 257 and 258:
239 [100] L. A. Smith, J. M. Bull,
Page 259 and 260:
241 [120] T. A. Welch. A technique
Page 261 and 262:
243 # Len Score Multicode % 1 30 10
Page 263 and 264:
Page 265 and 266:
Page 267 and 268:
Page 269 and 270:
Page 271 and 272:
253 Table A.5 continued: # Len Scor
Page 273 and 274:
Page 275 and 276:
Page 277 and 278:
259 the level of entropy within the
Page 279 and 280:
261 there are no pointers to these
Page 281 and 282:
263 information for each method so
Page 283 and 284:
265 to emit four files that were su
Page 285 and 286:
Figure B.1: The Steps Followed to E
Page 287 and 288:
269 B.7 Summary This Chapter has de
Page 289 and 290:
271 Selected Honors and Awards Cont
Page 291:
273 Departmental and University Pre
show all

OPTIMIZING THE JAVA VIRTUAL MACHINE INSTRUCTION SET BY ...

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?