Patent Application
20050021916
1. Technical Field
The present invention is directed to accesses to memory by input/output devices. More specifically, the present invention is directed to a system and method of improving fault-based multi-page pre-fetches.
2. Description of Related Art
Operating systems (OSs) as well as application programs are increasingly getting larger in size. Correspondingly, the amount of physical memory (i.e., random access memory or RAM) space they need to properly execute is also increasingly growing larger. However, to provide enough RAM space for simultaneous executions of an OS and an indeterminate number of application programs is unfeasible. Consequently, virtual memory is used.
Virtual memory is imaginary memory and is supported by most operating systems. (OSs and application programs will henceforth be referred to as programs.) Each executing program, which includes code and data, is allocated a certain amount of virtual memory space. For example in Windows-based systems, each executing program is allocated 2GB of virtual memory space. Thus, virtual memory is ostensibly limitless.
Each executing program is also allocated a certain amount of RAM space since a program must physically be in RAM in order to execute. However, allocated virtual memory space is usually much larger than allocated RAM space. Hence, a program that fits into its allocated virtual memory space may not all fit (especially its data) into its allocated RAM space. Since virtual memory does not really exist, the portion of the program that does not fit into the RAM is placed in a storage device (e.g., disk, tape, cartridge etc.). This allows the system to nonetheless execute the program by copying into RAM sections of the program needed at any given point during its execution.
To facilitate copying sections of a program into RAM, the operating system divides the virtual memory into virtual pages and the RAM into physical pages (also known as page frames). Each virtual page contains a fixed amount of space. Each physical page contains an equally fixed amount of space. Addresses of virtual pages are called virtual addresses and those of the physical pages are called physical addresses.
Thus, a page of data may either be in RAM or in a storage device. To keep track of which pages are in RAM, a virtual memory manager (VMM) is used. The VMM is a process that is primarily responsible for managing the use of both the RAM and the virtual memory. To do so, the VMM keeps track of each page that is in RAM and swaps pages to and from the storage device both on demand and routinely.
An on-demand swap occurs when an executing instruction of a program requests data that is not already in RAM. Particularly, when requesting data an executing program will provide the virtual address of the data. The virtual address is then translated into its physical address equivalent. If after the address translation the page on which the data is located is identified as being absent from the RAM, a page-fault exception is raised. A page fault results in switching immediately to a page fault handler. Using a replacement algorithm (e.g., a least recently used algorithm or LRU) a page of data in RAM is transferred onto the storage device. The page fault handler then loads the page onto which the requested data is located into the now-vacant page in RAM and, upon return, the instruction that generated the page fault is re-executed. This is a relatively fast process, but accumulating many page faults can have a drastic impact on performance.
Consequently, to reduce the number of page faults that may occur during the execution of a program, a method known as read-ahead or data pre-fetching is used. As the name suggests, data pre-fetching involves obtaining data before it is needed. Various types of data pre-fetching techniques have been developed. One of these techniques is called spatial data pre-fetching.
Spatial data pre-fetching is based on the likelihood that once data is referenced, nearby data is also likely to be referenced. That is, the decision to pre-fetch data is determined by the current data block access (e.g., fetching the data block adjacent to the data block currently being accessed).
Spatial data pre-fetching works splendidly when data is being read sequentially. For example, after two consecutive page faults of sequentially stored data, a block of sequential pages of data will be pre-fetched through normal data read-ahead. Hence, if future referenced pages are part of the pre-fetched block, which is highly likely when data is being read sequentially, the data will have already been in RAM when needed.
However, if data is being read randomly, spatial data pre-fetching may not work as well. For example, suppose an executing program is randomly reading data. Suppose further that the executing program makes a request to read a certain amount of data that resides on two sequential pages. If the data is not already in RAM, two page faults will be raised in order to load the two pages in the RAM. Because the pages are sequential, the system may infer that data is being read sequentially; and hence, pre-fetch a block of sequential pages of data. Since data is being read randomly, it is highly unlikely that future needed data will be on the pre-fetched block of pages. Thus, the block of pages may have been pre-fetched in vain and the physical pages onto which they are placed wasted. As will be explained later, continually pre-fetching unneeded pages of data may place an undue pressure on RAM space.
Thus a need exists for a system and method of improving multi-page fault-based data pre-fetches.
The present invention provides a system and method of improving fault-based multi-page pre-fetches. When a request to read data randomly from a file is received, a determination is made as to whether previous data has been read from RAM or from a storage device. If the data has been read from RAM, an attempt is made to read the present requested data from RAM. If the data is in RAM it is provided to the requester. If the data is not in RAM, a page fault occurs. If the requested data has a range that spans more than one page, the entire range is loaded in RAM by a page fault handler. If previous data has not been read from the RAM, it will be assumed that the present requested data is not in the RAM. Hence, the present requested data will be loaded into the RAM. Loading random data that spans a range of more than one page all at once into the RAM inhibits the system from pre-fetching data due to fault-based sequential data accesses.
In a particular embodiment, a trust value is assigned to a file when the file is opened. Each time data is to be read randomly from the file, the trust value is examined to determine whether previous data from the file was read from the RAM or from the storage device. If it is determined that previous data was read from the RAM it is assumed that the present requested data is also in the RAM. If the present requested data is indeed in the RAM, the trust value is incremented by a trust award. If the present requested data is not in the RAM, the trust value is decremented by a trust penalty. If, however, it is determined that previous data was not read from the RAM, it will be assumed that the present requested data is not in the RAM. The trust value is used to assist the system in determining whether the data is in the RAM or not. In any case, the data will be loaded into the RAM and the trust value will be incremented by the trust award.
The novel features believed characteristic of the invention are set forth in the appended claims. The invention itself, however, as well as a preferred mode of use, further objectives and advantages thereof, will best be understood by reference to the following detailed description of an illustrative embodiment when read in conjunction with the accompanying drawings, wherein:
With reference now to the figures,
The file system manager 102 interacts with executing programs. For example, when an executing program desires to read data, it provides the virtual address and the range of the data to the file system manager 102. The file system manager 102 will check with the VMM 112 to see whether the data is already in RAM (not shown). In so doing, the physical manager will provide the virtual address and the range of the data to the VMM 112.
The VMM 112 then translates the virtual address into a physical address. If after the translation it is determined that the data is in RAM, it is returned to the file system manager 102 which will pass it to the requesting program. If it is not in RAM, a page fault will be raised. Then, the VMM 112 will notify the file system manager 102 that the data is not in RAM. A page fault handler (not shown) will load the data from the storage device 134 into RAM. Once the data is loaded, the VMM 112 will notify the file system manager 102 that the data is now in RAM.
To retrieve the data from the storage device 134, the device driver 132 is contacted. A device driver, as is well known in the art, acts as a translator between a device and programs that use the device. That is, the device driver accepts generic commands from programs and translates them into specialized commands for the device.
As alluded to in the Description of the Related Art Section above, the VMM 112 anticipates future needs for pages of data of a file by observing the pattern used by a program that is accessing the file.
Thus, if a program is reading data randomly and in one instance requests a range of data that spans two, three, four etc. sequential pages that are not already in RAM, quite a number of pages may be pre-fetched in vain. If the program continually requests multiple pages of data randomly, the RAM may become over-committed. When the RAM is over-committed, thrashing may occur. Thrashing happens when the VMM continually pages data in and out of the RAM. When a system is thrashing, the system may not spend much time executing useful instructions; and thus, none of the active processes may make any significant progress. The present invention provides a heuristic algorithm that may be used to inhibit pre-fetching of data that is being accessed randomly.
The heuristic algorithm associates trust values with files. A trust value is a value that is used to enable a system to make an assumption as to whether or not requested data is in RAM. To do so, the invention starts by assigning a trust value of zero (0) to all files that are opened. Each time data is read from a file, its trust value is either incremented, if the assumption is correct, or decremented otherwise.
Specifically, when data is requested from a file, a determination (as is presently done) is made as to whether the file is being accessed sequentially or randomly. This determination is well known in the art and will not be explained. If the data in the file is being accessed sequentially, the invention will be bypassed. But, if the file is being accessed randomly, the file system manager will store the range of the read request in an in-memory inode.
To explain, each Unix directory entry contains a name and a pointer to an inode. The inode is associated with a file and includes the file size, permissions, a pointer to a sequence of disk blocks and one or two reference counts. A reference count is a number of file names that the file has. When a file is opened, its on-disk inode is read and converted into an in-memory inode. The in-memory inode is functionally identical to the on-disk inode except that it maintains a count of the number of processes that have opened the file in addition to the reference count maintained by the on-disk inode.
As mentioned above, the file system manager 102 will store the range of the data to be read in the in-memory inode of the file. For instance, if data with a range 8192, which spans two pages, is to be read, 8192 may be stored in the in-memory inode of the file. Then, the trust value associated with the file will be examined. If the trust value is greater than zero (trust value>0), it is an indication that previous data was read from RAM instead of from storage device 134. Therefore, it will be assumed that the present requested data is also in RAM. The trust value will then be increased by a value called a trust award and an attempt will be made to read the data from RAM.
If the data is in RAM, the data will be provided to the requesting program. If, however, the data is not in RAM, a page fault will occur. The VMM 112 will then notify the file system manager 102 and since the VMM does not keep the range of the data, it will also request that the file system manager 102 provide the range of the data again. The file system manager 102 will retrieve the range of the data from the in-memory inode of the file and provide it to the VMM 112. Using the range, the VMM 112 will load the data in RAM and the trust value will be decreased by a value called a trust penalty since the assumption was incorrect. Note that in this case the entire range of data will be read with just one page fault since the range is known. Hence, consecutive page faults will not be instituted in order to retrieve data stored on two or more consecutive pages and read-ahead will be obviated.
As explained above, the trust award is added to the trust value before the read is attempted. Thus, the trust penalty should be greater than the trust award otherwise the penalty will either be zero (0) or less. Further, a higher trust penalty allows for quick adaptability in the case where all future reads will generate a page fault. In this particular example, the trust award is one (1) and the trust penalty is twenty (20).
If the trust value is less or equal to zero (0), it is an indication that previously requested data was retrieved from the storage device 134 instead of from RAM. Therefore, it will be assumed that the present requested data will not be in RAM. Consequently, the file system manager 102 will pass the virtual address and the range of the data to be read to the VMM 112. The File system manager 102 will also instruct the VMM 112 to load the data in RAM. The VMM 112 will determine whether any of the pages of data to be read are already in RAM. All the pages that are not already in RAM will be retrieved from the storage device 134. Note that in this case pages will not be pre-fetched due to sequential data accesses because the data is not loaded in RAM as a result of a page fault; but rather, the data is pre-loaded into the RAM (i.e., before a fault occurs).
Here also, the trust value will be incremented by the trust award. In this case, the trust value is incremented to ensure that even if a file is not trusted, it will eventually be trusted after p reads, where p is less or equal to the trust penalty.
The invention should include a maximum trust value (i.e., the most a file will ever be trusted). This will allow the invention to adapt quickly to changing in-memory dynamics as for instance when a file that used to be in RAM and has been paged out onto the disk is being read. In that case, if the trust penalty is p and the trust maximum is m, the file will no longer be trusted after at most m/p faults, where m and p are integers. For example, a file which has a maximum trust value of 100 will no longer be trusted after 5 consecutive page faults if the trust penalty is 20. The minimum trust value however is p, the trust penalty, since when the trust value is less or equal to zero (trust value <=0), it will be incremented as each requested data will be assumed to be in the storage device 134.
The invention then provides an optimal way for a file system to adapt to a variety of random read workloads. These workloads include the case where data requested is entirely in the storage device 134 as well as the case where the data is wholly or partially cached. Indeed, the only suboptimal case occurs when a file is trusted (trust value >0) and the requested data fails to be in the RAM. In that case, a performance cost is paid for the attempt at reading the data from the RAM and for the resulting page fault. Nonetheless, the suboptimal case has performance advantages over present methods of reading data spanning more than one page into the RAM since only one page fault is used as opposed to the plurality of page faults that are ordinarily ensued.
If the file is being read randomly, the trust value of the file is examined to determine whether it is greater than zero. If the trust value is greater than zero, the range of data to be read is stored in the in-memory inode of the file and the trust value is incremented by a trust award (steps 306, 310, 312, 314 and 316). An attempt is then made to read the data from RAM (step 318). If the attempt is successful, the data is conveyed to the requesting program (steps 320 and 322). If the attempt is unsuccessful, the trust value will be decremented by a trust penalty and a page fault exception will be raised. The VMM will then obtain the range of data stored in the in-memory inode from the file system manager and will load in RAM all pages that are not already there. The data will then be provided to the requesting program before the process returns to step 304 (steps 320, 324, 326, 328, 330 and 332).
If the trust value is less or equal to zero (0), the trust value is incremented by the trust award and the process jumps to step 330 (steps 312 and 332). Note that once the process starts, it will stay running until the computer system on which it is implemented is turned off or all open files (e.g., their in-memory inodes) are uncached.
An operating system runs on processor 402 and is used to coordinate and provide control of various components within data processing system 400 in
Those of ordinary skill in the art will appreciate that the hardware in
The description of the present invention has been presented for purposes of illustration and description, and is not intended to be exhaustive or limited to the invention in the form disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art. Hence, the embodiment was chosen and described in order to best explain the principles of the invention, the practical application and to enable others of ordinary skill in the art to understand the invention for various embodiments with various modifications as are suited to the particular use contemplated.
