Profile-driven data layout optimization

• US 7,496,909 B2
• Filed: 02/28/2005
• Issued: 02/24/2009
• Est. Priority Date: 04/04/2000
• Status: Expired due to Fees

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Kuch, Gerald Dwayne, Zander, Jason L., Beckman, Brian C.
• Primary Examiner(s): Zhen; Wei Y
• Assistant Examiner(s): Nahar; Qamrun
• Application Number:
  US11/069,232
• Publication Number:
  US 20050177822A1
• Time in Patent Office: 1,457 Days
• Field of Search: 717/158, 717/151, 711/165
• US Class Current: 717/158
• CPC Class Codes: G06F 9/445 Program loading or initiating bootstrapping G06F9/4401; security arrangements for program loading or initiating G06F21/57
• CPC Class Codes: G06F 9/44557 Code layout in executable memory
• CPC Class Codes: G06F 9/4488 Object-oriented

Abstract

Data layout optimization arranges data members within memory to enhance software performance. Profiling data is consulted to determine how to group data members for an object class into groups. One technique groups the data members based on how frequently the data members are referenced in memory. Another technique groups the data members based on their affinities for one another in time as determined by observing when references to the data members take place. A variety of options when collecting the profiling data and grouping the data members is supported. The data member grouping is recorded in metadata associated with a definition of the object class. At runtime, a class loader places the data members of an object in memory according to the metadata. Data members of different groups can be placed in separately-loadable units of memory in the memory system. Subsequently, when the data members are referenced in memory, more frequently referenced data members, including those that tend to be referenced at times close to each other, reside at neighboring locations in the memory system.

Claims

1. One or more computer-readable media having computer-executable instructions for performing a method of splitting a class, the method comprising:

specifying at least one threshold corresponding to a certain number of memory references made to at least one of a plurality of data members of an object;

specifying a proposed cold group via the at least one threshold;

receiving an indication of a policy according to a policy framework supporting a plurality of options and configured to determine whether or not the class is to be split based at least on the proposed cold group, wherein the policy overrides a default policy by specifying the policy for a particular class, wherein the policy specifies at least one selected from the group consisting of;

the class is split if the proposed cold group is bigger than a hot group overhead only, the class is split if the proposed cold group is bigger than a hot group overhead plus the proposed cold group overhead, and the class is split if the proposed cold group is bigger than a hot group overhead plus a fraction of the proposed cold group overhead;

assigning memory locations for data members for a hot group of the object within a first unit of memory in a virtual memory system;

assigning memory locations for data members for the proposed cold group of the object within a second unit of memory in the virtual memory system separately loadable into primary memory from the first unit; and

splitting the plurality of data members of the object into at least the hot group and the proposed cold group based at least in part upon the at least one threshold and profile data corresponding to the plurality of data members of the object and responsive to determining that the policy indicates that the class is to be split.

2. The one or more computer-readable media of claim 1, further comprising using a threshold framework to allow for a plurality of ways to perform the specifying at least one threshold.

3. The one or more computer-readable media of claim 2, wherein one of the plurality of ways to perform the specifying at least one threshold comprises using an absolute threshold.

4. The one or more computer-readable media of claim 3, wherein the absolute threshold depends at least in part upon a total number of hits in a profiling run.

5. The one or more computer-readable media of claim 2, wherein one of the plurality of ways to perform the specifying at least one threshold comprises using a Gaussian threshold.

6. The one or more computer-readable media of claim 2, wherein one of the plurality of ways to perform the specifying at least one threshold comprises using a linear threshold.