Optimum write-back strategy for directory-based cache coherence protocols

• US 5,313,609 A
• Filed: 05/23/1991
• Issued: 05/17/1994
• Est. Priority Date: 05/23/1991
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Rathi, Bharat D., Baylor, Sandra J., McAuliffe, Kevin P.
• Primary Examiner(s): Robertson, David L.
• Application Number:
  US07/704,553
• Time in Patent Office: 1,090 Days
• Field of Search: 395/425
• US Class Current: 711/121
• CPC Class Codes: G06F 12/0817 using directory methods
• CPC Class Codes: G06F 12/082 Associative directories G06F12/0822 takes precedence

Abstract

A directory-based protocol is provided for maintaining data coherency in a multiprocessing (MP) system having a number of processors with associated write-back caches, a multistage interconnection network (MIN) leading to a shared memory, and a global directory associated with the main memory to keep track of state and control information of cache lines. Upon a request by a requesting cache for a cache line which has been exclusively modified by a source cache, two buffers are situated in the global directory to collectively intercept modified data words of the modified cache line during the write-back to memory. A modified word buffer is used to capture modified words within the modified cache line. Moreover, a line buffer stores an old cache line transferred from the memory, during the write back operation. Finally, the line buffer and the modified word buffer, together, provide the entire modified line to a requesting cache.

Claims

1. A system for optimizing the performance of a multiprocessor system having write-back caches and a multistage interconnection network connected to a shared memory, by allowing concurrent line transfer and write-back operations, comprising:

a modified word buffer configured to store any modified words within a modified cache line when said modified words are written back from a cache;

a line buffer configured to store an old cache line transferred from said memory, said old cache line corresponding in address identity with said modified cache line; and

a global directory associated with said modified word buffer and said line buffer, said global directory configured to store state and control information pertaining to cache lines, said global directory configured to request a write-back of said modified words from said cache and concurrently request said old cache line from said memory, and said global directory configured to cause said modified word buffer and said line buffer to write, in combination, said modified cache line to a requesting cache.

2. The system of claim 1, further comprising state lines available to said global directory corresponding to each cache line of the system, each of said state lines specifies a particular cache line, a global state, one or more processors owning the cache line, a wait event indicator indicating whether the state line is waiting for a write-back, and a request event indicator identifying the event which caused any write-back to occur.

3. The system of claim 1, further comprising a multistage interconnection network having three stages with four switches within each of said stages, said switches having two inputs and two outputs and configured to switch any input to any output.

4. A system for optimizing the performance of a multiprocessor system having write-back caches and a multistage interconnection network connected to a shared memory by allowing concurrent line transfer and write-back operations, comprising:

control means for receiving requests for a modified cache line from a requesting cache, for requesting a write-back of modified words in said modified cache line from a source cache and concurrently requesting a transfer of an old cache line from the memory, said old cache line corresponding in address identity with said modified cache line, and for transferring said modified cache line to said requesting cache; and

buffer means, connected to said control means, for capturing said modified words written back from said source cache and concurrently storing said old cache line from the memory.

5. The system of claim 4, further wherein and buffer means comprises a modified word buffer configured to store words and a line buffer configured to store said old cache line.

6. The system of claim 4, further comprising state lines available to said control means corresponding to each cache line of the system, each of said state lines specifies a particular cache line, a global state, one or more processors owning the cache line, a wait event indicator indicating whether the state line is waiting for a write-back, and a request event indicator identifying the event which caused any write-back to occur.

7. The system of claim 4, further comprising a multistage interconnection network having three stages with four switches within each of said stages, said switches having two inputs and two outputs and configured to switch any input to any output.

8. A method for controlling the performance of a multiprocessor system having write-back caches and a multistage interconnection network connected to a shared memory monitored by a global directory, the method for allowing concurrent line transfer and write-back operations, comprising the steps of:

(a) requesting the global directory for a modified cache line;

(b) requesting a write-back of modified words in said modified cache line from a source cache and concurrently requesting a transfer of an old cache line from the memory, said old cache line corresponding in address identity with said modified cache line;

(c) capturing in said global directory said modified words written back from said source cache and concurrently storing in said global directory said old cache line from said memory; and

(d) transferring from said global directory said modified cache line.

9. The method of claim 8, further including the step of identifying modified words by using a bit vector in a header.

10. The method of claim 8, further comprising the steps of determining at said global directory that data has been modified in said modified data line and determining the identity of said source cache.

11. The method of claim 8, further comprising the step of allocating a state line for each cache line of the system.

12. The method of claim 11, further comprising the step of specifying in said state line a particular cache line.

13. The method of claim 11, further comprising the step of specifying in said state line a global state.

14. The method of claim 11, further comprising the step of specifying in said state line one or more processors owning the cache line.

15. The method of claim 11, further comprising the step of providing a wait event indicator for indicating whether the state line is waiting for a write-back.

16. The method of claim 11, further comprising the step of providing a request event indicator for identifying the event which caused any write-back to occur.

17. The method of claim 8, further including the step of transferring said modified words simultaneously to both the memory and the global directory.