Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F9/00—Arrangements for program control, e.g. control units
  • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
  • G06F9/46—Multiprogramming arrangements
  • G06F9/466—Transaction processing
  • G06F9/467—Transactional memory

Definitions

• Transactional memory is a mechanism that allows concurrent programs to be written more simply.
• a transaction specifies a sequence of code that is supposed to execute "as if it were executing in isolation. In practice, transactions are allowed to execute concurrency, and conflicting data accesses are then detected. Decisions on what to do when contention occurs, i.e., "contention management," can have large effects on the performance of transaction systems.
• contention management can have large effects on the performance of transaction systems.
• contention management There is an important underlying problem in implementing a contention manager, however. Suppose that transaction TxI detects contention with transaction Tx2 - perhaps Tx2 holds a pessimistic write lock on a data item that TxI also wishes to lock.
• TxI In some situations, it might make sense for a contention manager to abort Tx2, allowing TxI to acquire the lock. Perhaps Tx2 is short, and TxI is long, so the cost of redoing Tx2's work after abort would be much less than redoing TxI 's.
• TxI noticed the contention, and thus seems like the logical place to execute the logic of the contention management decision. But to do so, it needs information about Tx2, such as statistics about its execution, the contents of its transaction log, or perhaps to request that Tx2 voluntarily aborts, freeing up its acquired locks. This information most naturally resides in the data structure representing the transaction Tx2. For efficiency reasons, it is desirable to have this data structure be local to the thread executing Tx2, rather than in some global data structure.
• the transactional memory system will define some locking data structure covering each possibly-shared data item. When a location is write- locked, this locking data structure will contain some indication of the transaction that holds the lock.
• TxI When TxI discovers that the data item it seeks to access is write-locked, it can read this locking data structure to discover the identity of the Tx2.
• Tx2 Once TxI had obtained a pointer to the data structure representing Tx2, and prepares to read information from that data structure, Tx2 may complete, and its data structure may be deallocated, and perhaps reallocated for some other purpose. In this situation, the information TxI reads about Tx2 is not stable.
• a transactional memory system provides a reference counting mechanism that enables transactions to safely examine data structures representing other transactions.
• the resolution of conflicts between two transactions (called "contention management") is facilitated using the reference counting mechanism. For example, because one transaction attempts to acquire an exclusive lock that another transaction owns, information about the owning transaction is obtained.
• a reference count of the data strucutre representing the owning transaction is incremented. The system ensures that the correct transaction data structure reference count was incremented, preventing that data structure from being deallocated and re-allocated to represent another transaction.
• the owning transaction data structure informs a contention management decision that determines whether to abort one of the two transactions or to wait for the owning transaction to release the lock.
• the contention management decision is made, the reference count of the owning transaction data structure is decremented by the conflicting transaction that previously incremented it.
• Each transaction data structure starts with a reference count of one, and, as each transaction completes, it decrements the reference count of its data structure.
• a non-zero reference count prevents the deallocation of a transaction data structure while a conflicting transaction in another thread is accessing the data structure. Data structures can be deallocated when their reference count is zero.
• dedicated type-stable allocation pools can be reduced using an unstable attribute.
• a thread unstable attribute is set before pointers to the transaction data structures can be acquired by threads, and cleared after uses of such pointers are complete.
• objects in a type-stable allocation pool can only be deleted if the thread unstable attribute is not set on any of the threads.
• Figure 1 is a diagrammatic view of a computer system of one implementation.
• Figure 2 is a diagrammatic view of a transactional memory application of one implementation operating on the computer system of Figure 1.
• Figure 3 is a high-level process flow diagram for one implementation of the system of Figure 1.
• Figure 4 is a process flow diagram for one implementation of the system of
• FIG. 1 illustrating the stages involved in making transaction log segments type- stable.
• Figure 5 is a process flow diagram for one implementation of the system of
• FIG. 1 illustrating the stages involved in making transaction data structures type- stable.
• Figure 6 is a process flow diagram for one implementation of the system of
• Figure 1 illustrating the actions taken by a transaction holding the lock that is participating in a reference counting mechanism.
• Figure 7 is a process flow diagram for one implementation of the system of Figure 1 that illustrates the actions taken by a transaction participating in a reference counting mechanism that notices contention.
• Figure 8 is a diagrammatic view of one implementation of using a reference counting mechanism between two transactions.
• Figure 9 is a process flow diagram for one implementation of the system of Figure 1 that illustrates the stages involved in reducing dedicated type-stable allocation pools using an unstable attribute.
• Figure 10 is a process flow diagram that illustrates an alternate implementation for reducing dedicated type-stable allocation pools by tracking location of pointers to type-stable memory.
• a transactional memory system uses type stability techniques to enable lock- free contention management.
• a reference counting mechanism is provided that allows one transaction to safely examine the data structure representing another transaction.
• the term "reference counting mechanism” as used herein is meant to include a technique for tracking a number or other data for a given transaction data structure that indicates if other transactions have an interest in the given data structure at a particular moment and for keeping the given transaction data structure from being deallocated (returned to the type-stable allocation pool) while other transactions have an interest.
• type-stable allocation pool denotes a pool of memory from which objects are allocated in a special way: once a block is allocated to represent an object of type T, that memory is never re-used to represent some other type U. Thus, a pointer to such a T may always be assumed to point to a T.
• safely examine as used herein is meant to include the ability to examine the data in a way that does not allow the data being examined to be deallocated while the examination is in process.
• a reference count of a transaction data structure is incremented by each other transaction that registers an interest in it. The reference count is decremented when this interest ends.
• Each transaction has "an interest” in the data structure that represents it, so these are allocated with reference count one, and when a transaction completes, the reference counts of its transaction data structure is decremented. Data structures cannot be deallocated until their reference count is zero.
• dedicated type-stable allocation pools can be freed safely through the use of an unstable attribute recognized by the garbage collector.
• a thread unstable attribute is set before pointers to the transaction data structures can be acquired by threads, and reset after all uses of these pointers have been completed.
• objects in a type-stable allocation pool can only be deleted if the thread unstable attribute is not set on any of the threads.
• an exemplary computer system to use for implementing one or more parts of the system includes a computing device, such as computing device 100.
• computing device 100 In its most basic configuration, computing device 100 typically includes at least one processing unit 102 and memory 104. Depending on the exact configuration and type of computing device, memory 104 may be volatile (such as RAM), non- volatile (such as ROM, flash memory, etc.) or some combination of the two. This most basic configuration is illustrated in Figure 1 by dashed line 106.
• device 100 may also have additional features/functionality. For example, device 100 may also include additional storage (removable and/or non-removable) including, but not limited to, magnetic or optical disks or tape.
• Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.
• Memory 104, removable storage 108 and nonremovable storage 110 are all examples of computer storage media.
• Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can accessed by device 100. Any such computer storage media may be part of device 100.
• Computing device 100 includes one or more communication connections 114 that allow computing device 100 to communicate with other computers/applications 115.
• Device 100 may also have input device(s) 112 such as keyboard, mouse, pen, voice input device, touch input device, etc.
• Output device(s) 111 such as a display, speakers, printer, etc. may also be included. These devices are well known in the art and need not be discussed at length here.
• computing device 100 includes transactional memory application 200. Transactional memory application 200 will be described in further detail in Figure 2.
• Transactional memory application 200 is one of the application programs that reside on computing device 100. However, it will be understood that transactional memory application 200 can alternatively or additionally be embodied as computer-executable instructions on one or more computers and/or in different variations than shown on Figure 1. Alternatively or additionally, one or more parts of transactional memory application 200 can be part of system memory 104, on other computers and/or applications 115, or other such variations as would occur to one in the computer software art.
• Transactional memory application 200 includes program logic 204, which is responsible for carrying out some or all of the techniques described herein.
• Program logic 204 includes logic for providing a contention manager 206; logic for making transaction log segments type-stable 208 (as described below with respect to Figure 4); logic for making transaction data structures type-stable 210 (as described below with respect to Figure 5); logic for using a reference counting mechanism to enable a first transaction to safely examine a state of a second transaction 212 (as described below with respect to Figures 6-8); logic for reducing dedicated type-stable allocation pools as appropriate 214 (as described below with respect to Figures 9-10); and other logic for operating the application 220.
• Figure 3 is a high level process flow diagram for transactional memory application 200. It should be noted that while Figure 3 is described in an order, there is no order intended, and the features and/or actions described in Figure 3 could be provided and/or performed in other orders.
• the process begins at start point 240 with making transaction log segments type-stable, as described in further detail in Figure 4 (stage 242). Type-stability guarantees that the object pointed by a type-stable pointer is of the indicated type.
• the process begins at start point 270 with ensuring that if a pointer is a pointer into a log segment, it will remain so at later times (stage 272).
• the system makes all log segments equal-sized, and aligned to their allocation size (stage 274).
• the system starts each log segment with a pointer to the data structure representing its owning transaction (stage 276). Given a pointer to a write log entry in a log segment, a pointer to the start of the log segment can be computed (stage 278).
• stage 278 Given a pointer to a write log entry in a log segment, a pointer to the start of the log segment can be computed (stage 278).
• stage 280 ends at end point 280.
• FIG. 5 illustrates one implementation of the stages involved in making transaction data structures type-stable.
• the process begins at start point 290 with providing a global pool of transaction data structures that only grows (stage 292).
• stage 292 To make allocation more efficient, medium sized chunks of them are given out to thread-local allocation pools (stage 294).
• stage 296 This makes a given pointer to a transaction data structure safe to dereference (stage 296).
• stage 296 This ends at end point 298.
• FIG. 6 illustrates one implementation of the actions taken by a transaction holding the lock that is participating in a reference counting mechanism.
• the process begins at start point 310 with giving a newly allocated transaction data structure a reference count of one, representing use by the transaction it represents (stage 312).
• stage 312 When each transaction completes, it decrements the reference count of its own transaction data structure (stage 314). If the reference count goes to zero, the data structure can be deallocated (stage 316). If reference count is greater than zero, it is not deallocated, but instead deallocation responsibility gets passed to the transaction that eventually reduces the reference count to zero (stage 318).
• stage 320 ends at end point 320.
• Figure 7 illustrates one implementation of the actions taken by a transaction participating in a reference counting mechanism that notices contention.
• the process begins at start point 340 with following locking information in an object to obtain a pointer to the data structure representing the possibly-owning transaction (stage 342).
• a reference count is incremented in the data structure of the possibly owning transaction (stage 344).
• determine if the reference count of the correct transaction data structure was incremented (decision point 346).
• the transaction data structure is for the "correct transaction” if the locking information still leads to the same transaction data structure.
• stage 348 If the correct reference count was not incremented, then the reference count of the data structure of the possibly owning transaction is decremented and locking is retried at a higher level (stage 348). [030] If the reference count of the correct transaction data structure was incremented (decision point 346), then a contention management decision is made on whether to abort the contending transaction, to abort self (the transaction that noticed the contention), or whether to wait for the contending transaction to release the lock (stage 350). The reference count of the owning transaction data structure is then decremented, and the data structure is deleted if the decrement takes its reference count to zero (stage 352). The process ends at end point 354.
• FIG 8 is a diagrammatic view of one implementation of using a reference counting mechanism between two transactions.
• the first transaction 362 has a data structure 364 and a transaction log segment 366.
• the second transaction 363 also has a data structure 365 and a transaction log segment 367.
• the second transaction 363 registers an interest in the data structure of the first transaction 362 because of a contended object 369
• the second transaction 363 reads locking information in the contended object 369 to discover that the object is locked, and that the first transaction 362 holds the lock.
• the locking information in contended object 369 contains a pointer to an entry in the log of the locking transaction 362, which contains more information about the locking.
• log segments contain pointers to their owning transaction data structure 364.
• the locking information might indicate the owning transaction directly, and other information about the locking stored in the log entry in log segement 366 might be kept in a hash table indexed by the address of the locked object.
• the second transaction increments the reference count (shown with a value of 2, after this increment) in the data structure 364 of the first transaction 362.
• the thread executing the first transaction is not stopped during this process. It may continue executing, complete its transactional work 362, and deallocate the data structure 364 that represented it. This data structure may even be re-allocated to represent a third transaction.
• the purpose of incrementing the reference count is to prevent deallocation, but the increment may occur after these steps, so that the data structure is deallocated or represents a different transaction.
• the second transaction 363 must therefore verify, after incrementing the reference count, that data structure 364 still represents the transaction of interest. It checks the locking information in contended object 364 again, verifying that it still leads to the same transaction data structure 364. If it does not, then the locking information of contended object 369 has changed; we therefore retry the locking process from the beginning, since it may now succeed. If it does lead to the same transaction data structure 364, transaction 363 has obtained a pointer 368 to the data structure 364 of the first transaction 362 so that it can safely examine the data structure 364 to facilitate a contention management decision for the contended object 369.
• the second transaction 363 When the second transaction 363 completes, it decrements the reference count of the data structure 364 in the first transaction 362 to indicate that it no longer has an interest in that data structure. A particular transaction data structure cannot be deallocated until its reference count is zero, which means that there are no longer any transactions interested in its data. If transaction 362 has previously completed, then the decrement by transaction 363 may take the reference count of transaction data structure 364 to zero, in which case transaction 363 must deallocate this data structure.
• Figure 9 illustrates one implementation of the stages involved in reducing dedicated type-stable allocation pools using an unstable attribute.
• the process begins at start point 370 with setting an unstable bit on the current thread (stage 372).
• a pointer is then acquired to a contending transaction (stage 374).
• each thread must set an unstable attribute before possibly acquiring pointers to transaction data structures.
• a contention management decision is made to resolve the conflict (stage 376). Once the contention management decision is made, the pointer to the contending transaction is no longer used (stage 376). Thus, the unstable bit is cleared on the current thread (stage 378).
• all threads are stopped at garbage collection, and no thread has its unstable attribute set, then objects in the type-stable allocation pools may be deallocated.
• Figure 10 illustrates an alternate to the implementation of Figure 9 that illustrates the stages involved in reducing dedicated type-stable allocation pools by tracking the location of pointers to type-stable memory.
• the process begins at start point 400 with requiring threads performing contention management to declare the location of pointers to type-stable memory (stage 402).
• the system performs a garbage-collection-like process on type-stable allocation pools, identifying the elements that are in use (stage 404).
• the system optionally deallocates some of the remaining type-stable allocation pools (stage 406).
• stage 406 The process ends at end point 408.

Landscapes

• Engineering & Computer Science (AREA)
• Software Systems (AREA)
• Theoretical Computer Science (AREA)
• Physics & Mathematics (AREA)
• General Engineering & Computer Science (AREA)
• General Physics & Mathematics (AREA)
• Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Priority Applications (3)

Applications Claiming Priority (2)

Publications (2)

Family

ID=40161866

Family Applications (1)

Country Status (6)

Cited By (1)

Families Citing this family (8)

Family Cites Families (26)

• 2007
  • 2007-06-29 US US11/824,353 patent/US7991967B2/en active Active
• 2008
  • 2008-06-10 TW TW097121596A patent/TWI455029B/zh not_active IP Right Cessation
  • 2008-06-10 TW TW103109965A patent/TWI515661B/zh not_active IP Right Cessation
  • 2008-06-18 EP EP08771367.3A patent/EP2176764B1/en not_active Not-in-force
  • 2008-06-18 JP JP2010514981A patent/JP5435741B2/ja not_active Expired - Fee Related
  • 2008-06-18 CN CN200880022472.XA patent/CN101689139B/zh not_active Expired - Fee Related
  • 2008-06-18 WO PCT/US2008/067346 patent/WO2009006025A2/en not_active Ceased
• 2011
  • 2011-08-02 US US13/196,569 patent/US8578105B2/en not_active Expired - Fee Related

Non-Patent Citations (2)

Also Published As

Similar Documents

Legal Events