Patent Grant
9678671
Technical Field
This disclosure relates to solid-state storage systems. In particular, this disclosure relates to a system and method for dynamically adjusting garbage collection policies in solid state storage systems.
Description of Related Art
Solid-state storage subsystems execute many commands in the course of their normal operation. For example, garbage collection is frequently performed on memory blocks that may contain both valid and invalid data. When such a memory block is selected for garbage collection, the garbage collection operation copies valid data within the memory block to a new location in memory and then erases the entire memory block, making the entire block available for future data writes. Therefore, the amount of memory freed by the garbage collection process depends on the amount of invalid pages within the memory blocks selected for garbage collection.
Systems and methods which embody the various features of the invention will now be described with reference to the following drawings, in which:
While certain embodiments of the inventions are described, these embodiments are presented by way of example only, and are not intended to limit the scope of the inventions. Indeed, the novel methods and systems described herein may be embodied in a variety of other forms. Furthermore, various omissions, substitutions and changes in the form of the methods and systems described herein may be made without departing from the spirit of the inventions.
Overview
Embodiments of the invention are directed to optimizing the selection of memory blocks for garbage collection to maximize the amount of memory freed by garbage collection operations. The systems and methods disclosed herein provide for the efficient selection of optimal or near-optimal garbage collection candidate blocks, with the most optimal selection defined as block(s) with the most invalid pages in one embodiment. In one embodiment, a controller classifies memory blocks into various invalid block pools by the amount (e.g., a minimum amount) of invalid pages each block contains. When garbage collection is performed, the controller in one embodiment selects a block from a non-empty pool of blocks with the highest minimum amount of invalid pages. The pools facilitate the optimal or near-optimal selection of garbage collection candidate blocks in an efficient manner and the data structure of the pools can be implemented with bitmasks, which take minimal space in memory.
One or more of the invalid block pools may have a dynamically adjustable minimum threshold of invalid pages for its blocks (e.g., a certain threshold percentage). In one embodiment, at least one such pool has a threshold amount I percentage that is set in accordance with an observed usage condition, which may reflect an the amount of over-provisioning (additional storage over stated capacity) provided in the non-volatile memory arrays, or a common range of percentage of invalid pages found in blocks that have been garbage collected. In one embodiment, as the amount of over-provisioning is adjusted, the minimum threshold of the dynamic invalid block pool is also adjusted. This optimizes the garbage collection process because there is a correlation between the amount of over-provisioning and the percentage of invalid pages that are likely to be found in the blocks, based on a mathematical property of the over-provisioned amount in relation to the stated capacity. Thus for example, a storage device with 25% over-provisioning may produce many blocks with 25% invalid pages once the storage device begins to reach into the provisioned amount. Setting the dynamic pool threshold to that amount can thus capture these blocks.
Although the present disclosure describes various embodiments as applicable to blocks, the embodiments are not so limited and are applicable to other units of memory such as superblocks. Also, as used in this application, “non-volatile memory” typically refers to solid-state memory such as NAND flask However, the systems and methods of this disclosure may also be useful in more conventional hard drives and hybrid drives including both solid-state and hard drive components. As such, while certain internal operations are referred to which typically are associated with solid-state drives, such as “wear leveling” and “garbage collection,” analogous operations for hard drives can also take advantage of this disclosure. Solid-state memory may comprise a wide variety of technologies, such as flash integrated circuits, Chalcogenide RAM (C-RAM), Phase Change Memory (PC-RAM or PRAM), Programmable Metallization Cell RAM (PMC-RAM or PMCm), Ovonic Unified Memory (OUM), Resistance RAM (RRAM), NAND memory, NOR memory, EEPROM, Ferroelectric Memory (FeRAM), or other discrete NVM (non-volatile memory) chips. The solid-state storage devices may be physically divided into planes, blocks, pages, and sectors, as is known in the art. Other forms of storage (e.g., battery backed-up volatile DRAM or SRAM devices, magnetic disk drives, etc.) may additionally or alternatively be used.
System Overview
The controller 150 also maintains several data structures including, in one embodiment, an invalid page table 152 and a mapping table 154. These data structures may reside in volatile memory such as DRAM. In one embodiment, the invalid page table 152 keeps track of the validity of data located at physical page addresses throughout the non-volatile solid-state memory arrays 160, while the mapping table 154 keeps track of the correspondence between logical block addresses (LBA) and physical page addresses in the non-volatile solid-state memory arrays 160.
In one embodiment, other data structures include invalid page counters 156, invalid block pools 162, and a free pool 164. In one embodiment, the assignment of blocks to the various pools are indicated by data markers such as bitmasks, though other methods of indication such as data flags or data tables are also possible. In one embodiment, invalid page counters 156 are maintained for at least some of the memory blocks in the memory arrays 160. In one embodiment, an invalid page counter 156 is maintained for each block and tracks the number of invalid pages within the associated block. In one embodiment, the invalid page counters 156 and/or the invalid page table 152 are stored in volatile memory such as dynamic random access memory (DRAM), with persistent copies stored in the non-volatile memory 160.
The memory blocks are assigned in one embodiment to various invalid block pools 162. For example, in a storage subsystem initially configured with 28% over-provisioning, blocks that are in use may be classified into a pool for blocks with at least 7% invalid pages, a dynamic pool for blocks with at least 28% invalid pages, a pool for blocks with at least 50% invalid pages, a pool for blocks with at least 75% invalid pages, or a pool for blocks with 100% invalid pages. Blocks that are available for new writes are classified into the free pool 164. As the amount of over-provisioning is adjusted, the dynamic pool may be adjusted accordingly. In one embodiment, the data structures and/or data related to the pool assignments are stored, for example, in volatile memory such as static random access memory (SRAM) and persistent copies may be additionally stored in the non-volatile memory 160. The use and maintenance of these data structures for garbage collection will be further described below.
Maintaining Invalid Page Data of Memory Blocks
For example, a write may be directed to LBA 10, which prior to the write operation was mapped to block 5, page 2. Before the actual write to memory, the controller updates the mapping table so that the entry for LBA 10 corresponds to the new location at which the data will be written (e.g., block 12, page 4). Alternatively, the controller may update the mapping table after the actual write to memory. In block 206, the controller determines whether an update to the invalid page table is needed. In most instances this is needed unless the particular LBA entry associated with the write operation has not been previously assigned to any physical address. In the above example, since block 5, page 2 is now invalid, the portion of the invalid page table 154 covering block 5 needs to be updated. If an update is needed, the method proceeds to block 208, where the controller increments the invalid page counter 156 for the block with the change. The invalid page table 152 is then updated in block 210.
The controller performs another check in block 212 to determine if the page counter for the block with the change has crossed an invalid page threshold. If a threshold has been crossed, the block is re-assigned to a new invalid block pool in block 214. Following the example above, if an invalid page counter indicates that the number of invalid pages within block 5 has increased to 1024 out of 2048 total, then block 5 is re-assigned from the 28% invalid pool to the 50% invalid pool.
In one embodiment, the controller process that handles write operations from the host obtains a lock on the invalid page counter, invalid page table, and the invalid block pool data structure or a subset of those data structures while performing the above referenced update tasks. Since the garbage collection process consults with some of the same data structures in its operation, locking these data structures prevents the garbage collection process and other internal system processes from reading outdated data and executing commands that would cause data consistency problems. For example, as the invalid page table is used in one embodiment to block pending garbage collection commands from being executed on physical page addresses indicated as containing invalid information, locking the table while the updating takes place ensures that garbage collection commands attempting to write invalid data are properly blocked from execution in accordance with updated information. In other embodiments the various update steps may be executed in a different order than that shown in
Using Invalid Page Data to Select Memory Blocks for Garbage Collection
The threshold percentages associated with the individual pools are provided in
The method 300 begins in block 302, where the controller determines whether there are enough blocks remaining in the free pool to accommodate pending write operations. Garbage collection is triggered if it is determined that there are not enough blocks in the free pool, and the method moves to block 304, where the controller determines if there are any blocks in the 100% invalid pool. If so, a memory block from that pool is selected in block 312. If not, the controller determines if there are any blocks in the 75% invalid pool in block 306. If so, a memory block from that pool is selected in block 312. The same process is repeated for the 50% invalid pool in block 308 if none are found in the 75% pool.
If there are no blocks in the 100%, 75%, or 50% pool, the method determines whether there is a block that can be chosen from the 28% (dynamic) pool in block 310. If not a block is taken from the 7% pool at block 314. The selected candidate block is then used in the garbage collection operation in block 316. In one embodiment, the 7% pool may set to be a fallback pool that matches the lowest (base) amount of over-provisioning in a storage device where the amount of over-provisioning is dynamically adjustable. The 7% may be based at least in part on an amount of difference between a 1,024 base and a 1,000 base storage unit accounting, as well as accounting for any other necessary storage space reserved to system operations, such as spare blocks allocated for replacing defective blocks. it is expected to contain at least one block because block 314 is reached only if the free pool is determined to have less than the threshold amount in block 302.
While the selection of a single block is shown in
Dynamic Invalid Block Pool
In block 354, the controller adjusts the dynamic block pool, For example, if the over-provisioning is changed from 25% to 46%, the dynamic block pool threshold may be changed from 25% invalid to 46% invalid as well. in other embodiments, the dynamic block pool is adjusted based on an observed usage condition, which may reflect an amount of over-provisioned capacity in non-volatile solid-state memory array, or may alternatively reflect a range of an amount of invalid pages that is frequently observed in memory blocks in non-volatile solid-state memory array. For example, the controller may observe that a large number of blocks are 22-33% and 90-93% invalid (through reviewing the invalid page counters of the blocks that have been garbage collected). The controller may adjust the one or more dynamic block pools to capture those blocks. For example, the controller may set one dynamic block pool to be a 22% minimum invalid pool and another dynamic block pool to be a 90% minimum invalid pool. In other embodiments, the controller may make such observation through a histogram which tracks the occurrence/frequency of blocks with certain percentages of invalid pages, and periodically adjust the dynamic pools according to recent histogram results. For example, a histogram with ten percentiles (e.g., 0% min. invalid, 10% min. invalid, 20% min. invalid, 30% min. invalid, etc.) may be used and one or more dynamic pools may be adjusted to match the percentile(s) with the most blocks.
In other embodiments where there are multiple dynamic block pools, some or all of the dynamic block pools may be adjusted as well. In one embodiment, the several dynamic block pools may have thresholds that are evenly distributed between 100% and X % invalid, where X is the amount of the current over-provisioning. Take, for example, the following setup for a 28% over-provisioning configuration: 100% (static), 76% (dynamic), 52% (dynamic), 28% (dynamic), and 7% (static). Since the difference between 100 and 28% is 72%, the 76% and 52% pools are spread out evenly within the difference span of 72% (two pools, 24% apart each). if the over-provisioning amount is changed to 40%, the pools may be updated to: 100% (static), 80% (dynamic), 60% (dynamic), 40% (dynamic), and 7% (static).
In block 356, the blocks affected by the changing of the dynamic pool thresholds are re-distributed. In one embodiment, this re-assignment of blocks to their appropriate pools may not be performed if the controller determines that the benefits do not outweigh the costs, since those affected blocks may eventually be garbage collected and removed from the pools. The costs of re-assignment involve performance costs associated with operations for updating the data to reflect the new pool assignments. On the other hand, post-reassignment, the benefits would be that each block is allocated to a more appropriate pool so garbage collection can recover more space per block and fewer data operations for garbage collection may be needed. Finally, in block 358, the garbage collection traversal order is updated to reflect the changes made to the pool(s). One such traversal order was previously described above in conjunction with
Multi-Pool Data Structure
Block Selection
In one embodiment, to achieve speed optimization, the selection of a block assigned to a pool with multiple blocks does not depend on the blocks' actual percentages of invalid pages and a candidate block within the pool may be selected based on the current location of a selection process or a block may be randomly selected from among blocks assigned to the same pool. The selection process may traverse the blocks in a consecutive fashion to locate a next available block assigned to the highest non-empty pool. For example, if the selection process last ended a search for a candidate block at block 1, block 5 would be selected since there are no more blocks in the 100% pool (block 7 has been taken) and block 5 is the first block in the 75% pool encountered in the selection process. Thus block 5 (90% invalid) may be selected from pool 4028 even though it has a lower percentage than block 0 (95% invalid) In one embodiment, the blocks may be selected by a round-robin fashion with one or more pointers traversing a list of blocks, locating a next block with the highest pool assignment. One pointer may be used per pool to indicate the last block taken for a particular pool. In other embodiments, additional checks and/or comparisons may be performed so a block with a higher or the highest percentage within the same pool may be selected. In various embodiments, several blocks are selected at once and may span across different pools. However, in some embodiments, the selection progression remains from the pool of blocks with the highest minimum amount of invalid pages to the pool of blocks with the lowest minimum amount of invalid pages.
Example Implementations
In one embodiment, the pool structure is implemented with bitmasks, and a block may be associated with one or more bits that indicate its assignment to the pools. For example, a four-pool structure may involve four corresponding bitmasks for the blocks, with each bitmask indicating whether the individual blocks belong to a particular pool. In one embodiment, additional checking bits may be assigned to a group of blocks to speed up the selection process. For example, a checking bit may be used to indicate the status of 32 blocks, such that when the checking bit is set to “0” the selection process can skip over the 32 blocks knowing that none of the blocks have a “1” bit indicating an assignment to the particular pool in question. If the checking bit is set to “1,” the selection process will check the individual bits for the 32 blocks since the “1” indicates that at least one block is assigned to the particular pool. In another embodiment, additional checking bits may be assigned to groups of checking bits in a hierarchal order. For example, an additional checking bit may be assigned to a group of checking bits so that if any of the checking bits is set to “1,”the additional bit will be set to “1” as well. The checking bit approach reduces the time needed to locate a block assigned to the pool with the highest minimum amount of invalid pages. In another embodiment, the pool assignments are maintained in a table.
The multi-pool data structure provides an efficient method for selecting optimal or near optimal candidate blocks for garbage collection. In one embodiment, a sorted linked list structure is used to organize the blocks that are eligible for garbage collection. in one embodiment, the blocks are sorted in the linked list by the amount of invalid pages in the blocks, so that the optimal candidate for garbage collection can be located by selecting a block from the front or back of the linked list (depending on the direction of the sort). In one embodiment, the above described pools are implemented in a sorted linked list structure with pointers to entries that correspond to the minimum invalid page thresholds of the individual pools, so that blocks assigned to the individual pools can be quickly located. In another embodiment, one or more linked lists are used for each pool, and blocks are assigned to a pool as described above and inserted into the one or more linked lists for the corresponding pool. In some embodiments, the pool data structure implemented with bitmasks may need substantially less overhead memory as compared to the linked list implementations. This difference can be substantial in storage subsystems in which there are potentially millions of blocks or tens of thousands of superblocks that may be candidates for garbage collection at any given time,
While certain embodiments of the inventions have been described, these embodiments have been presented by way of example only, and are not intended to limit the scope of the inventions. Indeed, the novel methods and systems described herein may be embodied in a variety of other forms. Furthermore, various omissions, substitutions and changes in the form of the methods and systems described herein may be made without departing from the spirit of the inventions. The accompanying claims and their equivalents are intended to cover such forms or modifications as would fall within the scope and spirit of the inventions. For example, those skilled in the art will appreciate that in various embodiments, the actual steps taken in the processes shown in
The present application claims the benefit of priority under 35 U.S.C. §120 as a continuation of U.S. patent application Ser. No. 13/173,266 entitled “System and Method for Dynamically Adjusting Garbage Collection Policies in Solid-State Memory,” filed on Jun. 30, 2011, now U.S. Pat. No. 9,158,670, the disclosure of which is hereby incorporated by reference in its entirety for all purposes.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5093780 | Sunahara | Mar 1992 | A |
| 5123094 | MacDougall | Jun 1992 | A |
| 5634099 | Andrews et al. | May 1997 | A |
| 5819310 | Vishlitzky et al. | Oct 1998 | A |
| 6067199 | Biumenau | May 2000 | A |
| 6237060 | Shilts et al. | May 2001 | B1 |
| 6338122 | Baumgartner et al. | Jan 2002 | B1 |
| 6430663 | Ding | Aug 2002 | B1 |
| 6475062 | Kubota et al. | Nov 2002 | B1 |
| 6493160 | Schreck | Dec 2002 | B1 |
| 6678801 | Greim et al. | Jan 2004 | B1 |
| 6711660 | Milne et al. | Mar 2004 | B1 |
| 6969989 | Mei | Nov 2005 | B1 |
| 7082494 | Thelin et al. | Jul 2006 | B1 |
| 7114041 | Hammitt et al. | Sep 2006 | B2 |
| 7146525 | Han et al. | Dec 2006 | B2 |
| 7174467 | Helms et al. | Feb 2007 | B1 |
| 7310699 | Sinclair | Dec 2007 | B2 |
| 7313641 | Koch et al. | Dec 2007 | B2 |
| 7315917 | Bennett et al. | Jan 2008 | B2 |
| 7340580 | Kinoshita | Mar 2008 | B2 |
| 7363396 | Liu et al. | Apr 2008 | B2 |
| 7363421 | Di Sena et al. | Apr 2008 | B2 |
| 7366826 | Gorobets et al. | Apr 2008 | B2 |
| 7424587 | Caulkins et al. | Sep 2008 | B2 |
| 7441054 | Wu et al. | Oct 2008 | B2 |
| 7441067 | Gorobets et al. | Oct 2008 | B2 |
| 7457903 | Purdham et al. | Nov 2008 | B2 |
| 7487286 | Nagano et al. | Feb 2009 | B2 |
| 7516355 | Noya et al. | Apr 2009 | B2 |
| 7582868 | Jiang et al. | Sep 2009 | B2 |
| RE41011 | Han et al. | Nov 2009 | E |
| 8521972 | Boyle et al. | Aug 2013 | B1 |
| 8725931 | Kang | May 2014 | B1 |
| 9026716 | Kang et al. | May 2015 | B2 |
| 20020083111 | Row et al. | Jun 2002 | A1 |
| 20020138694 | Isshiki | Sep 2002 | A1 |
| 20030051110 | Gaspard et al. | Mar 2003 | A1 |
| 20040179386 | Jun | Sep 2004 | A1 |
| 20040193743 | Byers et al. | Sep 2004 | A1 |
| 20050071537 | New et al. | Mar 2005 | A1 |
| 20050193081 | Gruber et al. | Sep 2005 | A1 |
| 20050216657 | Forrer et al. | Sep 2005 | A1 |
| 20060020849 | Kim | Jan 2006 | A1 |
| 20060041812 | Rajski et al. | Feb 2006 | A1 |
| 20060106981 | Khurshudov et al. | May 2006 | A1 |
| 20060155917 | Di Sena et al. | Jul 2006 | A1 |
| 20070016721 | Gay | Jan 2007 | A1 |
| 20070043900 | Yun | Feb 2007 | A1 |
| 20080082736 | Chow et al. | Apr 2008 | A1 |
| 20080091872 | Bennett et al. | Apr 2008 | A1 |
| 20080098192 | Im et al. | Apr 2008 | A1 |
| 20080235443 | Chow et al. | Sep 2008 | A1 |
| 20080263305 | Shu et al. | Oct 2008 | A1 |
| 20080288717 | Torabi | Nov 2008 | A1 |
| 20080307164 | Sinclair | Dec 2008 | A1 |
| 20090012976 | Kang et al. | Jan 2009 | A1 |
| 20090017220 | Muller et al. | Jan 2009 | A1 |
| 20090070502 | Noha et al. | Mar 2009 | A1 |
| 20090116475 | Krzyzanowski et al. | May 2009 | A1 |
| 20090150599 | Bennett | Jun 2009 | A1 |
| 20090154026 | Jiang et al. | Jun 2009 | A1 |
| 20090172258 | Olbrich et al. | Jul 2009 | A1 |
| 20090172260 | Olbrich et al. | Jul 2009 | A1 |
| 20090222618 | Cho | Sep 2009 | A1 |
| 20090313501 | Hallivuori | Dec 2009 | A1 |
| 20100088482 | Hinz | Apr 2010 | A1 |
| 20100325351 | Bennett | Dec 2010 | A1 |
| 20110010490 | Kwon et al. | Jan 2011 | A1 |
| 20110066788 | Eleftheriou et al. | Mar 2011 | A1 |
| 20110264843 | Haines et al. | Oct 2011 | A1 |
| 20110289260 | Wang et al. | Nov 2011 | A1 |
| 20120239851 | Calvert et al. | Sep 2012 | A1 |
| Entry |
|---|
| http://www.anandtech.com, Jun. 2011. |
| Taratorin, Alexander M. and Klaas B. Klassen, Observation of Recording Pole Instability in Perpendicular Recording, IEEE Transaction on Magnetics, vol. 42, No. 10, Oct. 2006, pp. 2267-2269. |
| Hu et al., “Write Amplification Analysis in Flash-Based Solid State Drives”, SYSTOR'09, Haifa, 2009, Zurich Research Laboratcny,copyright 2009 IBM Corporation, downloaded on May 12, 2010 from https://www.research.ibm.com/haifa/conferences/systor2009/papers/2—2—2.pdf, 13 pages. |
| Number | Date | Country | |
|---|---|---|---|
| 20160103617 A1 | Apr 2016 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 13173266 | Jun 2011 | US |
| Child | 14881103 | US |
