BRIEF DESCRIPTION OF DRAWINGS
So that the manner in which the above recited features, advantages and objects of the present invention are attained and can be understood in detail, a more particular description of the invention, briefly summarized above, may be had by reference to the embodiments thereof which are illustrated in the appended drawings.
It is to be noted, however, that the appended drawings illustrate only typical embodiments of this invention and are therefore not to be considered limiting of its scope, for the present invention may admit to other equally effective embodiments.
FIG. 1 is a sample Flash Based Drive architecture with multiple flash chips accessed by multiple Flash DMA engines according to an embodiment of the present invention.
FIG. 2 is a sample physical layout of data sections according to an embodiment of the present invention.
FIG. 3 is the LBA-PBA Map Table for the layout shown in FIG. 2 according to an embodiment of the present invention.
FIG. 4 is a physical layout with erased sections according to an embodiment of the present invention.
FIG. 5 is a block diagram illustrating how sections are placed to its new location when a write request is issued for that data section according to an embodiment of the present invention.
FIG. 6 is a flow chart illustrating the process of writing data to the Flash array according to an embodiment of the present invention.
FIG. 7 is a block diagram illustrating a list of pre-erased sections according to an embodiment of the present invention.
FIG. 8 is a block diagram illustrating the queue of pending operations for the flash DMA engines according to an embodiment of the present invention.
FIG. 9 is a block diagram illustrating how new write operations are to be added to the queue of the Flash DMA engines according to an embodiment of the present invention.
FIG. 10 is a block diagram illustrating an updated snapshot of the queue of pending operations according to an embodiment of the present invention.
FIG. 11 is a flowchart illustrating the process for Bad Block Management according to an embodiment of the present invention.
FIG. 12 is a sample physical layout with flash device level striping according to an embodiment of the present invention.
FIG. 13 is the LBA-PBA Map Table with striping for the layout shown in FIG. 12 according to an embodiment of the present invention.
DETAILED DESCRIPTION OF THE INVENTION
FIG. 1 shows an exemplary architecture that accommodates a very large number of flash arrays to achieve large capacities according to an embodiment of the present invention. The system comprises a number of Flash DMA Engines (FDEs) 101. A Flash DMA Engine (FDE) is basically an intelligent DMA controller that facilitates high speed data transfers to/from a group of flash memory devices. The system also contains a set of Flash Buses 102, which is a bus interface used by the FDE to connect to the flash memory devices. To increase capacity, a number of expansion boards 103 can be added. An expansion board is essentially a memory board that consists of a pool of flash memory devices for additional storage and a Flash Buffer Controller 104 for communication to the Flash DMA Engine. The Flash Buffer Controller is a controller that drives the Flash Bus and translates the command signals from the FDEs into native flash commands that can be understood by the target flash chip. The number of buses/engines can be increased/decreased according to the required performance, cost, and storage capacity of the system.
The flash array organization comprises a set of Flash DMA engines controlling multiple flash devices across a set of Flash Buses. The set of flash devices assigned to a particular Flash Bus is called a “flash array bank”. Each bank can be partitioned into any number of flash array banks with the Flash DMA engines sharing a Flash Bus. For example in FIG. 1, it is shown that a group of n number of Flash DMA engines such as 105 is sharing a single Flash Bus0106.
Each flash DMA engines is assigned to control a set of flash devices. This set of flash devices is said to belong to a flash array bank interleave. In addition, each flash device within this interleave is said to belong to a different flash group. From the figure, all flash chips labeled ‘A0’ within flash array bank interleave 107 is controlled by Flash DMA Engine A0 and each of the flash device within this interleave belongs to a different group, i.e. first flash device A0108 belongs to Group 0, second Flash Device A0109 belongs to Group 1, and so on.
To optimize access to this very large array of flash devices, a number of operations are done in parallel. There are three methods of interleaving that are easily supported in this system; these are bus, flash array bank and group interleaving. Bus interleaving is the transfer of data to/from flash devices using the different Flash Buses. The flash array bank interleaving method on the other hand, is the transfer of data to/from flash devices belonging to the same bus but in different Flash DMA engines. Lastly, group interleaving is the transfer of data by a certain Flash DMA engine to/from different flash devices it controls.
The main advantage of implementing the bus interleaving method is that the flash access is done in parallel utilizing the different Flash Buses, i.e. Flash Bus 0, Flash Bus 1, Flash Bus 2, and so on. Each Flash DMA engine uses a different Flash Bus in order to achieve parallel operations. Flash array bank interleaving has parallel operations during flash access by utilizing the busy signal status of the active Flash Bus. As an example, one engine (FDE A0, where FDE stands for flash DMA engine, the term FDE and DMA engine is used interchangeably in this document) is writing data to a flash device (Flash Device A0) while FDE A0 is waiting for the command completion, other FDE of different bank interleave, e.g., FDE A1, can access Flash Bus 0 and send out a command to a different target flash device such as flash device A1. Accordingly, group interleaving performs parallel operations by having a specific FDE send multiple commands to different flash devices it controls. As an example, one engine (FDE A0) sends a command to a flash device A0 of Group0. While FDE A0 is waiting for the command to be completed and the Flash Bus is temporarily idle, FDE A0 can send another command to a flash device in another group, e.g., flash device A0 of Group1, in order to achieve optimum data transfer.
From this, it can be seen that data transfers are most efficient if flash devices are accessed using different flash bus (bus interleaving), then using different Flash DMA engine (flash array bank interleaving) and lastly different group (group interleaving). Another feature of new flash devices is its multi-bank capability. A single flash device is sub-divided into 4 banks wherein parallel operation can occur. In a Multi-Bank operation, an FDE can target up to 4 different blocks in a target flash device and up to 4 blocks can be erased and/or programmed using a single request.
To take advantage of this parallel operation, a mapping scheme that considers all these capabilities must be created. To lessen the size of the LBA-PBA Map Table, a section size is defined to be the minimum relocatable area. Assuming an LBA size is 512 bytes and the section size is 4 Kbytes, only 1 entry is needed for every 8 LBAs. The section size is primarily limited by the native page size of a flash device (a page is smaller then the minimum erase size or flash block, a flash block is made up of multiple pages). It is always a multiple of this page size since a NAND flash is usually programmed on a per page basis. Since the section size is the minimum relocatable region, when only 5 LBAs are updated, the other 3 LBAs must be relocated together with the new data. Smaller section would therefore lead to more flexibility but larger overhead to maintain the LBA-PBA mapping. Although large section might suffer because of the need to relocate the unmodified data, typical OS usually accesses the media in larger blocks like 4 KB. The choice of the section size depends largely on how the host accesses the media. The larger the host access is, the more acceptable it is to use large section size to minimize the LBA-PBA mapping without suffering from the need to relocate unmodified data. Taking the concept wherein applications for rotating drives tend to optimize sequential access, this system as illustrated in FIG. 1 should take advantage of this and optimize for sequential access. Therefore, an exemplary ideal layout is illustrated in FIG. 2.
FIG. 2 is a sample physical layout of data sections according to an embodiment of the present invention. For illustrative purposes, the system shown in FIG. 2 has 16 DMA engines with 2 engines sharing a bus. Each engine also controls two flash devices for a total of 32 flash devices. In this example, a section consists of 8 LBAs. As can be seen from FIG. 2, consecutive sections are distributed all throughout the entire flash arrays taking advantage of bus interleaves, then engine interleaves, then group interleaves. In this way, when the hosts requests for LBAs 0-23, equivalent to sections 0-2201 (24 LBAs is equivalent to 3 sections), the requests will go through FDEs 0, 2, 4202 utilizing buses 0, 1, 2203. This layout is ideal for sequential access. This layout is ideal because the mapping is able to take advantage of bus interleaving (then bank interleaving and group interleaving) that the system provides. So whenever the host accesses data sequentially, the mapping assures that the system will fetch the data in the most efficient or parallel way, taking advantage of the bus interleaving then bank interleaving and then group interleaving. But as noted before, due to the inherent characteristic of flash devices requiring erase cycles before writing new data, write operations will trigger the data to be relocated to new locations that have been previously erased (to save on erase cycles).
FIG. 3 is the LBA-PBA Map Table for the layout shown in FIG. 2 according to an embodiment of the present invention. FIG. 3 shows how the LBA-PBA Map Table will look like based on the FIG. 2 layout. A section consists of a group of LBAs. In this example, a section (corresponding to a row in the table) consists of 8 LBA. The PBA stored here contains the information for both the location of the Flash Device, uniquely identified using its Engine Number (Bus Number was added to illustrate bus interleaving but each Engine is associated with only one Bus) and Group Number, and the address within the flash device. From FIG. 2, the first physical section of Dev0 is labeled as Section 0, the first physical section of Dev1 is labeled as Section1 . . . , the second physical section of Dev0 is labeled as Section 32, and so on. Correspondingly, in FIG. 3, Section 0301 is located at Dev 0, which has a unique address Bus 0, Engine 0, Group 0 and is in address 0 within that flash device. Section 1302 is located at Dev 1, which has a unique address Bus 0, Engine 0, Group 0 and is in address 0 within that flash device. Section 61303 is located at Dev 27, which has a unique address Bus 5, Engine 11, Group 1 and is in address 0x08 within that flash device. Assuming the flash is addressable every 512 bytes and a section is 4 Kbytes in size, address 0x08 represents the second physical section (or the second 4 Kbyte unit) within a flash device, address 0x10 the third physical section (or the third 4 Kbyte unit) and so on. The 512 byte addressable unit means that every flash address represents 512 bytes so address 0x00 is the first 512 bytes, address 0x01 the second 512 bytes, and so on. That leads to address 0x00-0x07 representing first 4 Kbytes of a flash device and 0x08-0x0F the next 4 Kbytes. The 512 byte addressable unit is just an arbitrary value for the system, it can be byte addressable leading to address 0x0000-0x0FFF representing the first 4 Kbytes and address 0x1000-0x1FFF the next 4 Kbytes.
Mapping also plays a major role to look up target physical locations for bad block management, for wear-leveling and most importantly for write operations. In write operations, instead of writing the new data in its old physical location, an erased physical location is obtained and the logical block is remapped there to save an erased cycle. Determining the new physical location is dependent on the current load of the system. As mentioned, the illustrated storage system works in the most optimum way when it takes advantage of the parallel operations it can execute at a given time with the Bus interleaving being the most beneficial (then Engine interleaving, then Group interleaving). That means that whenever the system needs to determine a new physical location, it must take this into consideration. When the system currently uses FDE 0 utilizing Bus0 and FDE 2 utilizing Bus 1, it would prioritize looking for physical locations located in Flash Devices that have an address with Bus Number not equal to 0 or 1 to take advantage of Bus interleaving. On the other hand, if the system has already utilized Buses 0-7, it then checks what particular engine that's not being used to take advantage of engine interleaving (i.e. if Engine 0, 1, 2, 3, 4, 6, 8, 10, 12, 14 is in use, look for a location that's addressed in either Engine 5, 7, 9, 11, 13, 15 because that's then one that can take advantage of engine interleaving).
FIG. 4 is a physical layout with erased sections according to an embodiment of the present invention. FIG. 4 shows the distributed erased sections of the system. These erased sections will be the ones used when a write request need a new location. Only a small amount of the total storage needs to be reserved, this area can be used in conjunction with bad block replacements. The reserved amount depends on the capacity of the drive and the frequency of write requests it is subjected to. For a 60 GB drive, a 2-3% area or about 1-2 GB is efficient enough to provide erased sections as demanded with the stale blocks erased in the background. The location of the free sections needs only to be distributed to all the flash arrays in the system and can be distributed in anyway within the flash device since access time within a flash is unaffected by its location. However, for flash devices offering multi-bank support, a flash device is divided into different banks with concurrent operation capability. For this case, the reserved pre-erased sections are distributed to the different banks of a flash chip to take advantage of the concurrent operation feature.
To prevent inefficient access to the storage system due to the need to relocate data to a different physical location on every write request, the placement policy assures that the resources are utilized properly. FIG. 5 is a block diagram illustrating how sections are placed to its new location when a write request is issued for that data section according to an embodiment of the present invention. Due to the write request for LBAs 0-23, sections 0-2, which holds this data, was relocated from its previous location 501 to the previously erased sections 502. This would mean that when the host reads this data, it is still fetched in parallel as before. So not only will this approach increase efficiency during the write operation, it was also able to provide efficient access for future read operations of this data. For flash-based systems, the important thing is to spread out the data to different flash chips to take advantage of parallel access. Initially, the drive itself is optimized for sequential access (as seen in FIG. 2) but as the host issues the write requests, the flash based system quickly adopts to the current load. The placement policy would force all the write requests to spread evenly across the entire system. As a consequence of spreading the locations, all consecutive requests would naturally go to different flash chips. In a typical scenario wherein related requests are done together, future reads to this data would also be optimized since related blocks would probably be spread out. How optimized the consequent layout would be largely dependent on how the Host issues its write requests. In a single process, this would be most optimized since requests are being generated by one source only. In a multi-user/process environment, requests come from different sources and the Host typically interleaves this. But even so, the placement policy would still be able to spread out related blocks although not as efficient since it is interleaved with other requests.
FIG. 6 shows the typical process when doing writes to flash devices. Basically whenever the system needs to write data to the flash array, it first determines the current load of the FDE to see where to put the new data. In addition to determining the FDE, it can also determine to which flash device controlled by that FDE is optimum. If for example, it will put the request to FDE 8 and sees there is a request for Group 0, it will then prioritize the placement to Group 1 to take advantage of multi-group operation. To take advantage of the multi-bank support of new flash devices, target sections can also consider this i.e. if there's a request for Flash Dev 4 targeting Bank 0, it can prioritize a request for Bank 1 of that flash device. This is of course after bus, engine, group prioritization.
FIG. 7 shows a simple way of listing all the available erased sections that can be utilized per engine for easy fetching of new locations. It holds the physical location 701, which contains all the needed information like the FDE number, Group location, and block address within the chip. Additionally, it contains the section index 702, which is the next free section within a flash block. For simplicity and because of certain flash chips limitation that pages within a flash block can be programmed sequentially only, this index will just increment from 0 (entire flash block is erased) to the max index (3 if there are 4 sections in a block i.e. 4K section in a 16K block). So if only one section is needed, the index will just be incremented but if an entire block is needed, only the free entry with Section 0 can be utilized.
FIG. 8 shows a sample snapshot of the queue operations waiting for each Flash DMA engine. When a new write request is needed, the placement policy will determine what physical locations are prioritized. The easiest way is just to do a round-robin scheme wherein the priority just rotates in a fixed manner (FDE 0, 2, 4, . . . 14, 1, 3 . . . 15 for Group 0 flash devices then for Group 1, etc). This assures that all write operations have been evenly distributed. The drawback on this is that it doesn't consider the other operations that the engines are currently working on and thus might cause some unwanted delays. This can be resolved by using some threshold in which if a certain engine has too much workload with respect to others, that engine will simply be skipped in the rotation (i.e. if a particular engine has 5 more entries than the minimum of all the engines, it will just be skipped in the rotation and the next will be given the workload).
FIG. 9 shows how the first two write requests being added to the queue of flash DMA engine operations. It shows the fetching of free sections from their respective engines and using that section for the new write requests. Assuming no request has been finished, the queue may look like FIG. 10 after queuing up 20 write requests evenly distributed through all DMA engines.
FIG. 11 shows the flow on how to manage bad blocks. When the system detects a bad block, it needs to remap it to a good block. To retain the distribution created when that data was first written, the target block will be prioritized by looking at a free block (from the erased list with Section=0) on the same flash device, then different flash device controlled by the same engine, flash devices on a different engine but the same bus and so on.
To further spread out the user data without increasing the size of the LBA-PBA Map Table, a striping feature can be utilized. Striping forces parallel access at a lower level compared to what a section size can provide. There are programmable parameters for striping, one is the stripe size and the other is the number of chip interleaves. Stripe size means the number of contiguous LBAs in a chip before moving on to the next chip. Number of chip interleaves means the number of chips the stripe will be distributed.
FIG. 12 shows the layout of FIG. 2 with striping support. Keeping section 1201 size to be 4 KB, stripe 1202 size is 1 KB and the number of chip interleaves 1203 is 4. This means 32 LBAs will be distributed to 4 chips (8 per section) with 2 LBAs comprising a stripe. FIG. 13 shows how the map table would look like. There would be 4 related entries for a set of 32 LBAs or 4 sections. Size of the map table would still be the same but the LBAs will be striped to the different chips in each set. LBA 0, 1 would be in Dev 0; LBA 2, 3 in Dev 1; LBA 4, 5 in Dev 2; LBA 6, 7 in Dev 3; LBA 8, 9 back to Dev 0 and so on. As a consequence, when the host requests a read to LBAs 0-7. Engines 0, 2, 4, and 8 are activated to get the 8 LBAs instead of just engine 0 when there's no striping.
This is mostly advantageous for hosts that only access the disk one request at the time and those requests are small chunks of data (i.e. request only 4 KB and waits for a response every time). Without striping, the host was only able to utilize a few of the available resources. This advantage decreases once the host is capable of queuing up multiple requests or requests large amount of data. This is because in this scenario, the host requests have already forced the different flash DMA engines to work in parallel and thus utilized the system resources efficiently.
In one embodiment of the present invention, an apparatus for data storage comprises: a plurality of flash buses; a plurality of DMA engines coupled to at lease two of the plurality of flash buses; and a plurality of flash chips coupled to at least two of the plurality of DMA engines; wherein data access performance is improved by bus interleaving wherein one or more data is transferred to or from the plurality of flash chips using at least two flash buses; wherein data access performance is improved by flash array bank interleaving wherein one or more data is transferred to or from the plurality of flash chips using at least two DMA engines; and wherein data access performance is improved by group interleaving wherein one or more data is transferred to or from the plurality of flash chips using at least two flash chips. Optionally, each of the plurality of flash chips further comprises a plurality of sections; each section in the apparatus is operable to be accessed using a physical block address comprising a least significant portion, a second least significant portion, a third least significant portion, and a fourth least significant portion; wherein the least significant portion comprises an order according to the plurality of flash buses, the second least significant portion comprises an order according to a plurality of DMA engines each coupled to a same flash bus, the third least significant portion comprises an order according to a plurality of flash chips each coupled to a same DMA engine, and the fourth least significant portion comprises an order according to the plurality of sections in a same flash chip; and wherein a logical block address for host data access is mapped to a physical block address according to a placement algorithm whereby host data access performance is improved. Optionally, at least one section in at least one of the plurality of flash chips is designated as free section; at least one free section is pre-erased as target for at least one data relocation whereby flash write performance and bad block replacement performance are improved according to the placement algorithm. Optionally, the placement algorithm maps each logical block address to a physical block address in a linear mapping prior to the at least one data relocation; and the placement algorithm designates a plurality of free sections evenly to each of the plurality of flash chips. Optionally, the placement algorithm locates a target for each of the at least one data relocation for write operation according to a current load represented in a system queue based on a priority comprising firstly locating a free section corresponding to a different flash bus, secondly locating a free section corresponding to a different DMA engine coupled to a same flash bus, and thirdly locating a free section corresponding to a different flash chip coupled to a same DMA engine; and the placement algorithm locates a target for each of the at least one data relocation for bad block management based on a priority comprising firstly locating a free section corresponding to a same flash chip, secondly locating a free section corresponding to a different flash chip coupled to a same DMA engine, and thirdly locating a section corresponding to a different DMA engine coupled to a same flash bus. Optionally, each section further comprises a plurality of strips; and the linear mapping further comprises one or more chip interleaves.
Foregoing described embodiments of the invention are provided as illustrations and descriptions. They are not intended to limit the invention to precise form described. In particular, it is contemplated that functional implementation of invention described herein may be implemented equivalently in hardware, software, firmware, and/or other available functional components or building blocks, and that networks may be wired, wireless, or a combination of wired and wireless. Other variations and embodiments are possible in light of above teachings, and it is thus intended that the scope of invention not be limited by this Detailed Description, but rather by Claims following.
Patent Application
20070288686
