CA2126754A1 - Method for performing disk array operations using a nonuniform stripe size mapping scheme - Google Patents

Abstract

METHOD FOR PERFORMING DISK ARRAY OPERATIONS
USING A NONUNIFORM STRIPE SIZE MAPPING DEVICE

ABSTRACT OF THE DISCLOSURE
A method and apparatus for improving disk performance in a disk array subsystem. A nonuniform mapping scheme is used wherein the disk array includes regions having varying sizes of data stripes. The disk array includes a region comprised of data stripes having a stripe size that corresponds to the size of the internal data structures frequently used by the file system, in addition to a region comprised of a number of data stripes having a larger stripe size which are used for general data storage. When a write operation occurs involving one of the data structures, the data structure is preferably mapped to the small stripe region in the disk array having a size which matches the size of the data structure. In this manner, whenever a file system data structure is updated, the operation is a full stripe write. This removes the performance penalty associated with partial stripe write operations.

Description

~ - -- 21~ 6 1~ ~

METHOD FOR PERFORMING DISK ARRAY OPERATIONS
VSING A NONUNIFORM STRIPE SIZE MAPPING SCHE~E

The present invention is directed toward a method for improving perf~rmance for multiple disk drives in computer systems, and more particularly to a method for performing write operations in a disk array utilizing parity data redundancy and recovery protection.

Nicroprocessors and the computers which utilize -them have become increasingly more powerful during the recent years. Currently available personal computers have capabilities in excess of the mainframe and -minicomputers of ten years ago. Microprocessor data bus sizes of 32 bits are widely available whereas in the past 8 bits was conventional and 16 bits was common.
Personal computer systems have developed over t~e years and new uses are being discovered daily. The uses are varied and, as a result, have different ~- ~
requirements for various subsystems forming a complete ;
computer system. With the increased performance of computer system~, it became apparent that mass storage subsystems, such as fixed~disk drives, played an increasingly important role in the transfer of data to and from the computer system. In the past few years, a -new trend in storage subsystems, referred to as a disk array subsystem, has emerged for improving data transfer performance, capacity and reliability. One reason for building a disk array subsystem is to create 21267~ -

-2-a logical device that has a very high data transfer rate. This may be accomplished by "ganging" multiple standard disk drives together and transferring data to or from these drives in parallel. Accordingly, data is stored "across" each of the disks comprising the disk array so that each disk holds a portion of the data comprising a data file. If n drives are ganged together, then the effective data transfer rate may be increased up to n times. This technique, known as striping, originated in the supercomputing environment where the transfer of large amounts of data to and from secondary storage is a frequent requirement. In striping, a sequential data block is broken into segments of a unit length, such as sector size, and sequential segments are written to sequential disk drives, not to sequential locations on a single disk drive. The unit length or amount of data that is stored "across" each disk is referred to as the stripe size. The stripe size affects data transfer characteristics and access times and is generally chosen to optimize data transfers to and from the disk array. If the data block is longer than n unit lengths, the process repeats for the next stripe location on the respective disk drives. With this approach, the n physical drives become a single logical device and may be implemented either through software or hardware.
One technique that is used to provide for data protection and recovery in disk array subsystems is referred to as a parity scheme. In a parity scheme, data blocks being written to various drives within the array are used and a known EXCLUSIVE-OR (XOR) technique is used to create parity information which is written to a resèrved or parity drive within the array. The advantage of a parity scheme is that it may be used to 67~

minimize the amount of data storage dedicated to data redundancy and recovery purposes within the array. For example, Figure 1 illustrates a traditional 3+1 mapping scheme wherein three disks, disk 0, disk 1 and disk 2, are used for data storage, and one disk, disk 3, is ~
used to store parity information. In Figure 1, each -rectangle enclosing a number or the letter "p" coupled with a number corresponds to a sector, which is preferably 512 bytes. As shown in Figure 1, each -complete stripe uses four sectors from each of disks O, 1 and 2 for a total of 12 sectors of data storage per disk. ~Assuming a standard sector size of 512 bytes, the stripe size of each of these disk stripes, which is defined as the amount of storage allocated to a stripe -on one of the disks comprising the stripe, is 2 kbytes.
Thus each complete stripe, which includes the total of the portion of each of the disks allocated to a stripe, can store 6 kbytes of data. Disk 3 of each of the stripes is u~ed to store parity information. However, 2~ there are a number of disadvantages to the use of parity fault tolerance techniques in disk array systems. One disadvantage to the 3+1 mapping scheme illustrated in Figure 1 is the loss of performance within the disk array as the parity drive must be updated each time a data drive is updated or written ;~
to. The data must undergo the XOR process in order to write the updated parity information to the parity drive as well as writing the data to the data drives.
This process may be partially alleviated by having the parity data also distributed, relieving the load on the dedicated parity disk. However, this would not reduce the number of overall dzta write operations.
Another disadvantage to parity fault tolerance techniques is that traditional operating systems perform many small writes to the disk subsystem which ~ -~

- 212675~

.

are often smaller than the stripe size of the disk array, referred to as partial stripe write operations.
For example, many traditional file systems use small data structures to represent the structure of the files, directories, and free space within the file system. In a typical ~NIX file system, this information is kept in a structure called an INODE, which is generally 2 kbytes in size. The INODE or INDEX NODE contains information on the type and size of the file, where the data is located, owning and using users, the last time the file was modified or accessed, the last time the NODE was modified, and the number of links or file names associated with that INODE. In the OS/2 high performance file system, this structure is called an FNODE. These structures are updated often since they contain file access and modification dates and file size information. These structures are relatively small compared with typical data stripe sizes used in disk arrays, thus resulting in a large number of partial stripe write operations.
When a large number of partial stripe write operations occur in a disk array, the performance of the disk subsystem is seriously impacted because, as explained below, the data or parity information currently on the disk must be read off of the disk in order to generate the new parity information. This results in extra revolutions of the disk drive and causes delays in servicing the request. In addition to the time required to perform the actual operations, it will be appreciated that a READ operation followed by a WRITE operation to the same sector on a disk results in the loss of one disk revolution, or approximately 16.5 milliseconds for certain types of hard disk drives.
Where a complete stripe of data is being written to the array, the parity information may be generated 2~,~67~

-5~
directly from the data being written to the drive array, and therefore no extra read of the disk stripe is required. However, as mentioned above, a problem occurs when the computer writes only a partial stripe to the disk array because the disk array controller -does not have sufficient information from the data to be written to compute parity for the complete stripe.
Thus, partial stripe write operations generally require data stored on a disk to first be read, modified by the process active on the host system, and written back to the same address on the data disk. This operation -consists of a data disk READ, modification of the data, and a data disk WRITE to the same address. There are -~
generally two techniques used to compute parity -~;
information for partial stripe write operations.
In the first technique, a partial stripe write to a data disk in an XOR parity fault tolerant system includes issuing a READ command in order to maintain parity fault tolerance. The computer system first reads the parity information from the parity disk for the data disk sectors which are being updated and the old data values that are to be replaced from the data disk. The XOR parity information is then recalculated by the host or a local processor, or dedicated logic, by XORing the old data sectors to be replaced with the related parity sectors. This recovers the parity value without those data values. The new data values are XORed on to this recovered value to produce the new parity data. A WRITE command is then executed, writing the updated data to the data disks and the new parity information to the parity disk. It will be appreciated ~ -~
that this process requires two additional partial sector READ operations, one from the parity disk and one readinq the old data, prior to the generation of the new XOR parity information. Additionally, the ~ ~
~ ~ :

212&7~
--' WRITE operations are to locations which have just been read. Consequently, data transfer performance suffers.
The second method requires reading the remainder of the data that is not to be repudiated for the stripe, despite the fact that it is not being replaced by the WRITE operation. Using the new data and the old data which has been retrieved, the new parity information may be determined for the entire stripe which is bein~ updated. This process requires a READ
operation of the data not to be replaced and a full stripe WRITE operation to save the parity information.
According to the prior art, a disk array utilizing parity fault tolerance had to perform one of the above technigues to manage partial stripe WRITE operations.
Therefore, partial stripe writes hurt system performance because either the remainder of the stripe that is not being written must be fetched or the existing parity information for the stripe must be read prior to the actual write of the information.
Accordingly, there exists a need for an improved method for performing disk WRITE operations in a parity fault tolerant disk array in order to decrease the number of partial stripe write operations.
Background on disk drive formatting is deeme~
appropriate. When a disk drive is produced or manufactured, the manufacturer will also generally have low level formatted the disk. A low level format operation involves the creation of sectors on the disk along with their address markings, which are used to identify the sectors after the formatting is completed.
The data portion of the sector is established and filled in with dummy data. When a disk drive unit is incorporated into a computer system, the disk controller and the respective operating system in control of the computer system must perform a high ~ - 2~26754 :~:
- ' :

level or logical format of the disk drive to place the "file system" on the disk and make the disk drive conform to the standards of the operating system. This high level formatting is performed by the respective disk controller in conjunction with an operating system service referred to as a "make file system" program.
In the UNIX operating system, the make file system program works in conjunction with the disk controller to create the file system on the disk array. In traditional systems, the operating system views the disk as a sequential list of blocks or sectors, and the make file system program is unaware as to the top~logy ~ -of these blocks.

The present invention is directed toward a method and apparatus for improving disk performance in a computer system having a disk array subsystem. In the method according to the present invention, a nonuniform mapping scheme is used wherein the disk array includes certain designated regions having varying sizes of data stripes. The disk array includes a region comprised of a number of data stripes having a stripe size that is approximately the same as the size of internal data structures frequently used by the file system, in addition to a region which includes a number of data stripes having a larger stripe size which are used for general data storage. When a write operation occurs involving one of the small data structures, the data ~-structure is preferably mapped to the small stripe region in the disk array wherein the complete stripe size matches the size of the data structure. In this manner, whenever the file system data structure is updated, the operation is a full stripe write. This reduces the number of partial stripe write operations, ~1~67~

thus reducing the performance penalty associated with these operations.

A better understanding of the present invention can be obtained when the following detailed description of the preferred embodiment is considered in conjunction with the following drawings, in which:
Figure 1 is a prior art diagram of a traditional

3~1 disk array mapping scheme having a uniform stripe size;
Figures 2 and 3*are block diagrams of an illustrative computer system on which the method of the present invention may be practiced; (*Figs. 3A and 3B);
lS Figure 4 is a block diagram of the disk subsystem of the preferred embodiment;
Figure S is a functional block diagram of the transfer controller of Fig. 4 according to the : :~
preferred embodiment;
Figure 6 is a diagram of a 3+1 disk array mapping ~ ~:
scheme having varying stripe sizes according to a first embodiment; ~ : :
Figure 7 is a diagram of a RAID 5 3+1 disk array : -mapping scheme having varying stripe sizes according to -~
a second embodiment of the invention; :
Figure 8 is a diagram of a 4+1 disk array mapping .
scheme according to the preferred embodiment of the invention;
Figure 9 is a flowchart diagram of a WRITE
operation according to the method of the present invention; and Figure lO is a flowchart diagram of a READ
operation according to the method of the present invention.

''` '~'`'''''' :~ ''-~ . .' ` 2~2675~

g The computer system and disk array subsystem described below represent the preEerred embodiment of the present invention. It is also contemplated that -other computer systems, not having the capabilities of the system described below, may be used to practice the present invention.
Referring now to Figs. 2 and 3, the letter C
generally designates a computer system on which the present invention may be practiced. For clarity, system C is shown in two portions with the interconnections between Figs. 2 and 3 designated by reference to the circled numbers 1 to 8. System C is comprised of a number of block elements interconnected via four buses.
A central processing unit CPU comprises a system processor 20, a numerical co-processor 22, a cache memory controller 24, and associated logic circuits connected to a system processor bus 26. Associated with cache controller 24 is a high speed cache data random access memory (RAM) 28, non-cacheable memory address (NCA) map programming logic circuitry 30, non~
cacheable address or NCA memory map 32, address ~
exchange latch circuitry 34, data exchange transceiver - -36 and page hit detect logic 43. Associated with the CPU also are system processor ready logic circuit 38, next address (NA) enable logic circuit 40 and bus request logic circuit 42.
The system processor is preferably an Intel Corporation 80386 microprocessor. The system processor 20 has its control, address and data lines interfaced ~ ~ -to the system processor bus 26. The co-processor 22 is ;
preferably an Intel 80387 and/or Weitek WTL3167 numerical processor interfacing with the local processor bus 26 and the system processor 20 in the 2~ 267~

conventional manner. The cache RAM 28 is preferably a suitable high-speed static random access memory which interfaces with the address and data elements of bus 26 under the control of the cache controller 24 to carry out required cache memory operations. The cache controller 24 is preferably an Intel 82385 cache controller configured to operate in two-way set associative master mode. In the preferred embodiment, the components are the 33 MHz versions of the respective units. An Intel 80486 microprocessor and an external cache memory system may replace the 80386, numeric coprocessor, 8238S and cache RAM i~ desired.
Address latch circuitry 34 and data transceiver 36 interface the cache controller 34 with the processor 20 and provide a local bus interface between the processor bus 26 and a host or memory bus 44. Circuit 38 is a logic circuit which provides a bus ready signal to control access to the bus 26 and indicate when the next cycle may begin. The enable circuit 40 is utilized to indicate that the next address of data or code to be utilized by sub-system elements in pipelined address -~
mode may be placed on the local bus 26.
Non-cacheable memory address ~NCA) map programmer 30 cooperates with the processor 20 and the non~
cacheable address memory 32 to map non-cacheable memory locations. The non-cacheable address memory 32 is utilized to designate areas of the system memory that are non-cacheable to avoid various types of cache coherency problems. The bus request logic circuit 42 is utilized by the processor 20 and associated elements to request access to the host bus 44 in situations 6uch as when requested data is not located in cache memory 28 and access to system memory is required.
The main memory array or system memory 58 is coupled to the host bus 44. The main memory array 58 r~ ~
212~7~'~

is preferably dynamic random access memory. Memory 58 interfaces with the host bus 44 via EISA bus buffer (EBB) data buffer circuit 60, a memory controller circuit 62, and a memory mapper 68. The buffer 60 performs data transceiving and parity generating and checking functions. The memory controller 62 and memory mapper 68 interface with the memory 58 via address multiplexor and column address strobe (ADDR/CAS) buffers 66 and row address strobe (RAS) enable logic circuit 64.
In the drawings, Sysitem C is configured as having the processor bus 26, the host bus 44, an extended industry standard architecture (EISA) bus 46 (Fig. 3) and an X bus 90 (Fig. 3). The details of the portions of the system illustrated in Fig. 3 and not discussed in detail below are not significant to the present invention other than to illustrate an example of a fully configured computer system. The portion of System C illustrated in Fig. 3 is essentially a configured EISA system which includes the necessary EISA bus 46 and EISA bus controller 48, data latches and transceivers referred to as EBB data buffers 50 and ~ -address latches and buffers 52 to interface between the EISA bus 46 and the host bus 44. Also illustrated in ~
Fig. 2 is an integrated system peripheral (ISP) 54, ~ --which incorporates a number of the elements used in an ~ -EISA-based computer system.
The integrated ISP 54 includes a direct memory access controller 56 for controlling access to main memory 58 (Fig. 1) or memory contained in an EISA slot ~ :
and input/output (I/0) locations without the need for access to the processor 20. The ISP 54 also includes interrupt controllers 70, non-maskable interrupt logic 72, and system timer 74 which allow control of interrupt signals and generate necessary timing signals 21267~

and wait states in a manner according to the EISA
specification and conventional practice. In the preferred embodiment, processor generated interrupt requests are controlled via dual interrupt controller circuits emulating and extending conventional Intel 8259 interrupt controllers. The ISP 54 also includes bus arbitration logic 75 which, in cooperation with the bus controller 48, controls and arbitrates among the various requests for EISA bus 46 by the cache controller 24, the DMA controller 56, and bus master devices located on the EISA bus 46.
The EISA bus 46 includes ISA and EISA control buses 76 and 78, ISA and EISA data buses 80 and 82, and are interfaced via the X bus 90 in combination with the ISA control bus 76 from the EISA bus 46. Control and data/address transfer for the X bus 90 are facili~ated by X bus control logic 92, data buffers 94 and address buffers 96.
Attached to the X bus are various peripheral devices such as keyboard/mouse controller 98 which interfaces with the X bus 90 with a suitable keyboard and a mouse via connectors 100 and 102, respectively.
Also attached to the X bus are read only memory (ROM) circuits 106 which contain bas~c operation software for the system C and for system video operations. A serial port communications port 108 is also connected to the system C via the X bus 90. Floppy disk support, a parallel port, a second serial port, and video support circuits are provided in block circuit 110.
The computer system C includes a disk subsystem 111 which includes a disk array controller 112, fixed ~ -disk connector 114, and fixed disk array 116. The disk array controller 112 is connected to the EISA bus 46, preferably in a slot, to provide for the communication of data and address information through the EISA bus -~ 21,~7~

46. Fixed disk connector 114 is connected to the disk array controller 112 and is in turn connected to the fixed disk array 116.
Referring now to Fig. 4, the disk subsystem 111 used to illustrate the method of the present invention is shown. The disk array controller 112 has a local processor 130, preferably an Intel 80186. The local processor 130 has a multiplexed addresstdata bus UAD
and control outputs UC. The multiplexed address/data bus UAD is connected to a transceiver 132 whose output is the local processor data bus UD. The multiplexed address/data bus UAD is also connected to the D inputs -of a latch 134 whose Q outputs form the local processor address bus UA. The local processor 130 has associated with it random access memory (RAM) 136 coupled via the multiplexed address/data bus UAD and the address data bus UA. The RAM 136 is connected to the processor control bus UC to develop proper timing signals. -~
Similarly, read only memory (ROM) 138 is connected to the multiplexed address/data bus UAD, the processor address bus UA and the processor control bus UC. Thus, the local processor 130 has its own resident memory to control its operation and for its data storage. A
programmable array logic (PAL) device 140 is connected -to the local processor control bus UC to develop --additional control signals utilized in the disk array controller 112.
The local processor address bus UA, the local processor data bus, UD and the local processor control bus UC are also connected to a bus master interface controller (BMIC) 142. The BMIC 142 serves the function of interfacing the disk array controller 112 with a standard bus, such as the EISA or MCA bus, and acts as a bus master. In the preferred embodiment, the BMIC 142 is interfaced with the EISA bus 46 and is the .

2~ 2~7~
` -- , Intel 82355. Thus, by this connection with the local processor busse~ UA, UD and VC, the BMIC 142 can interface with the local processor 130 to allow data and control information to be passed between the host system C and the local processor 130.
Additionally, the local processor data bus UD and local processor control bus UC are preferably connected to a transfer controller 144. The transfer controller 144 is generally a specialized multi-channel direct ~
memory access (DMA) controller used to transfer data -: -between the transfer buffer RAM 146 and various other devices present in the disk array controller 112. For example, the transfer controller 144 is connected to the BMIC 142 by the BMIC data lines BD and the BMIC
control lines BC. Thus, over this interface, the transfer controller 144 can transfer data from the transfer buffer RAM 146 to the BMIC 142 if a READ
operation is requested. If a WRITE operation is requested, data can be transferred from the BMIC 142 to the transfer buffer RAM 146. The transfer controller 144 can then pass this information from the transfer buffer RAM 146 to disk array 116. The transfer ~ :
controller 144 is described in greater detail in U.S.
Application No. 431,735, and in its European counterpart, European Patent Office Publication No.
0427119, published April 4, 1991, which is hereby incorporated by reference. ~:
The transfer controller 144 includes a disk data bus DD and a disk address bus and control bus DAC. The disk address and control bus DAC is connected to two buffers 165 and 166 which are part of the fixed disk connector 114 and are used to send and receive control signals between the transfer controller 144 and the disk array 116. The disk data bus DD is connected to two data transceivers 148 and 150 which are part of the :~

21~7~l~

fixed disk connector 114. The outputs of the transceiver 148 and the transfer buffer 146 are connected to two disk drive port connectors 152 and 154. In similar fashion, two connectors 160 and 162 are connected to the outputs of the transceiver 150 and the buffer 166. Two hard disks can be connected to ~-each connector 152, 154, 160, and 162. ~hus, up to 8 -~
disk drives can be connected and coupled to the transfer controller 144. As discussed below, in the preferred emb3diment five disk drives are coupled to the transfer controller 144, and a 4+1 mapping scheme is used.
In the illustrative disk array system 112, a compatibility port controller (CPC) 164 is also connected to the EISA bus 46. The CPC 164 is connected to the transfer controller 144 over the compatibilit~
data lines CD and the compatibility control lines CC.
The CPC 164 is provided so that the software which was written for previous computer systems which do not have a disk array controller 112 and its BMIC 142, which are addressed over an EISA specific space and allow very high throughputs, can operate wit~out requiring a rewriting of the software. Thus, the CPC 164 emulates the various control ports previously utilized in interfacing with hard disks.
Referring now to Fig. 5, the transfer controller 144 is itself comprised of a series of separate circuitry blocks. The transfer controller 144 includes two main units referred to as the RAM controller 170 and the disk controller 172. The RAM controller 170 has an arbiter to control the various interface devices that have access to the transfer buffer RAM 146 and a multiplexor so that the data can be passed to and from the transfer buffer RAM 146. Likewise, the disX
controller 172 includPs an arbiter to determine which - 2~f~7~

of the various devices has access to the integrated disk interface 174 and includes multiplexing capability ~ .
to allow data to be properly transferred back and forth :~
through the integrated disk interface 174.
The transfer controller 144 preferably includes 7 DMA channels. One DMA channel 176 is assigned to cooperate with the BMIC 142. A second DMA channel 178 is designed to cooperate with the CPc 164. These two devices, the BMIC 142 and the bus compatibility port controller 164, are coupled only to the transfer buffer RAM 146 through their appropriate DMA channels 176 and 178 and the RAM controller 170. The BMIC 142 and the compatibility port controller 164 do not have direct access to the integrated disk interface 174 and the disk array 116. The local processor 130 (Fig. 3) is connected to the RAM controller 170 through a local processor DMA channel 180 and is connected to the disk controller 172 through a local processor disk channel 182. Thus, the local processor 130 is connected to both the transfer buffer RAM 146 and the disk array 116 as desired.
Additionally, the transfer controller 144 includes~

4 DMA disk channels 184, 186, 188 and 190 which allow~- .
. information to be independently and simultaneously passed between the disk array A and the RAM 146. It is noted that the fourth DMA/disk channel 190 also ---~
includes XOR capability so that parity operations can be readily performed in the transfer controller 144 :
without requiring computations by the local processor~
30 - 130. The above computer system C and disk array subsystem 111 represent the preferred computer system for the practice of the method of the present invention.
The computer system C preferably utilizes the UNIX ~:
operating system, although other operating systems may .,:~ :.,: .

7 ~

be used. As described in the background, the UNIX
operating system includes a service referred to as the make file system program. In the preferred embodiment, the ~ake file system program provides information to the disk controller 112 as to how many INOD~s are being created and the size of the INODEs. optionally, the make file system includes sufficient intelligence to inform the disk controller 112 as to the desired stripe size in the small stripe and large stripe regions and the boundary separating these regions. As previously discussed, the number of INODEs is approximately equal tc the number of files which are to be allowed in the system. The disk array controller 112 uses this information to develop the file system on each of the disks comprising the array 116.
The disk array controller 112 uses a multiple mapping scheme according to the present invention which partitions the disk array 116 into small stripe and large stripe regions. The small stripe region preferably occupies the first N sectors of each disk and is reserved for the INODE data structures, and the remaininq stripes in the array form the large stripe region, which comprises free space used for data storage. Therefore, in the preferred embodiment, the disk controller 112 allocates the first N sectors of each of the disks in the array for the small stripe region. The remaining sectors of each of the disks are formatted into the large stripe region. The disk array controller 112 stores the boundary separating the smal~l stripe and large stripe regions in the RAM 136.
Thereafter, when the INODE data structures are written to each of the disks, the disk array controller 112 utilizes this boundary and writes the INODEs to the small stripe portion of the array 116. In this manner, whenever an INODE is updated, the resulting operation `

-` 2~2~7~

is a full stripe write. This increases system performance because, as previously discussed, partial stripe write operati~ns reduce performance because they generally require a preceding read operation. In an alternate embodiment of the invention, the small stripe region does not occupy the first N sectors of each disk, but rather the small stripe region includes a plurality of regions interspersed among the large stripe regions. In this embodiment, a plurality of boundaries which separate the small stripe and large stripe regions are stored in the RAM 136 so that the disk controller 112 can write the INODE data structures to the small stripe region.
In an alternate embodiment of the invention, the ;~
OS/2 operating system is used. In this embodiment, an ; -OS/2 service similar t~ the make file system program discussed above provides information to the disk controller 112 as to how many FNODEs are being created and the size of the FNODEs. The disk controller then uses a multiple mapping scheme similar to that discussed above to partition the disk array 116 into small stripe and large stripe regions wherein the small stripe region is reserved for the FNODE data ~-structures. It is noted that the present invention can -operate in conjunction with any type of operating system or file system.
Referring again to Figure 1, when an INODE data structure having a size of 2 kbytes is written to a disk array having a uniform disk stripe size of 2 kbytes according to the prior art, for example stripe ~ ~-0, then the entire INODE would be written to disk 0 in sectors 0, 1, 2 and 3, and disks 1 and 2 would be unused. Effectively, this operation negates the advantages of a disk array system since only one disk is being accessed. Also, the amount of unused space is `- 2~2~75~

considerable, resulting in an inefficient use of the disk drives. If data is subsequently allowed to be written to the remainder of the stripe, i.e., the portion of the stripe residing in disks 1 and 2, then the resulting ~peration in this stripe will consist of a partial stripe write operation, which has performance penalties as described above.
Referring now to Figure 6, a diagram illustrating a 3+1 mapping scheme utilizing multiple stripe sizes according to one embodiment of the present invention is shown. Figure 6 is exemplary only, it being noted that the disk array 116 will utilize a much larger number of stripes of each size. The disk drives used in the preferred embodiment include a number of sectors each having 512 bytes of storage. In the embodiment shown in Figure 6, the disk array utilizes two stripe sizes, a disk stripe size of one kbyte and a disk stripe size of two kbytes. As shown in Figure 6, stripes 0, 1 and 2 utilize a disk stripe size of one k~yte using two sectors per disk in disks 0, 1 and 2 for a total of six sectors or 3 kbytes of data storage per complete stripe. In addition, stripes 0, 1, and 2 utilize two sectors in disk 3 to store parity information for each stripe. Stripes 3-6 utilize a 2 kbyte disk stripe size wherein four sectors per disk on disks 0, 1 and 2 are allocated for data storage for each stripe and four sectors on disk 3 are reserved for parity information for each stripe. In this embodiment, the INODE data structures are written to the portion of the disk array having the small stripe size, i.e., stripes 0, 1, or 2.
As previously discussed, INODE structures are assumed to be 2 kbytes in size in the preferred embodiment.
Therefore, as shown in Figure 6, INODE structures written to stripes 0, 1 or 2 would fill up the area in disks 0 and 1, disk 2 would generally be unused, and 212~5'~ -disk 3 would be used to store the respective parity information. In this embodiment, data would not be allowed to be written to disk 2 of the respective stripe after an INODE is written there, and thus partial stripe write operations are prevented from occurring. Therefore, by using a smaller stripe size in a portion of the disk array 116 and preventing data from being written to the unused space after an INODE
is written, a write operation of these structures emulates a full stripe write. However, it is noted that disk 2 is unused or unwritten during this full stripe write, and thus an inefficient use of the disk area results. In addition, since disk 2 will generally be unused for each small stripe where an INODE
structure is written, the data transfer bandwidth from the disk array system is reduced, and the array essentially operates as a 2+1 mapping scheme in these instances.
One solution to this problem is to distribute the un~sed space across different disks for each stripe as shown in Figure 7. In this manner, the reduction of data transfer bandwidth is not as significant since each disk is used approximately equally. However, data transfer bandwidth is sub-optimum since each disk access involves an unused disk. Furthermore, this method produces an undesirable amount of unused disk space.
In the preferred embodiment of the invention, a 4+1 mapping scheme is used, as shown in Figure 8. It is again noted that Figure 8 is exemplary only, and the disk array 116 of the preferred embodiment will utilize a much larger number of stripes in each of the small stripe and large stripe regions. The disk stripe size of the stripes in the small stripe region, stripes 0-4, wherein stripe size is defined as the amount of each 2~2~7~l~

disk that is allocated to the stripe, is 512 bytes. In this manner, each complete stripe in the small stripe region holds exactly 2 kbytes of data, which is approximately equivalent to the size of an INODE
structure. Accordingly, when an INODE structure is written to a small stripe in the disk array 116, a full stripe write operation is performed, and the INODE
occupies the entire stripe without any unused space.
In this manner, the bandwidth of the disk array 116 is optimally used because every disk participates in each access and no unused space results.
Referring again to Figure 4, in the preferred embodiment, a disk request is preferably submitted by the system processor 20 to the disk array controller 112 through the EISA bus 46 and BNIC 142. The local processor 130, on receiving this request through the BMIC 142, builds a data structure in the local processor RAM memory 136. This data structure is known as a command list and may be a simple READ or WRITE
request directed to the disk array 116, or it may be a more elaborate set of requests containing multiple READ/WRITE or diagnostic and configuration requests.
The command list is then submitted to the local processor 130 for processing. The local processor 130 then oversees the execution of the command list, including the transferring of data. Once the execution of the command list is completed, the local process~r 130 notifies the operating system device driver running on the system microprocessor 20. The submission of the command list and the notification of the command list completion are achieved by a protocol which uses input/output (I/O) registers located in the ~MIC 142.
The READ and WRITE operations executed by the disk array controller 112 are implemented as a number of application tasks running on the local processor 130.

21~ 6 7 5 ~ ::

Because of the nature of the interactive input/output operations, it is impractical for the illustrative computer system C to process disk commands as single -batch tasks on the local processor 130. Accordingly, the local processor 130 utilizes a real time multi-tasking use system which permits multiple tasks to be addressed by the local processor 130, including the method of the present invention. Preferably, the operating system on the local processor 130 is the -AMX86 multi-tasking executive by Kadak Products, Ltd.
The AMX operating system kernel provides a number of system services in addition to the applications set forth in the method of the present invention.
Referring now to Figure 9, a flowchart diagram of a WRITE operation as carried out on a computer system C
including the intelligent disk array controller 112 is ~ --shown. The WRITE operation begins at step 200, in which the active process or application causes the system processor 20 to generate a WRITE request which is passed to the disk device driver. The disk device driver is a portion of the software contained within the computer system C, preferably the system memory 58, which performs the actual interface operations with the disk units. The disk device driver software assumes control of the system processor 20 to perform specific tasks to carry out the required I/O operations.
Control transfers to step 202, wherein the disk device driver assumes control of the system processor 20 and generates a WRITE command list.
In step 204, the device driver submits the WRITE
command list to the disk controller 112 via the BMIC
142 or the CPC 164. The device driver then goes into a wait state to await a completion signal from the disk array controller 112. Logical flow of the operations proceeds to step 206, wherein the local processor 130;~
.

- 2~2675~ :

receives the WRITE command list and determines whether an INODE data structure is being written to the disk array 116. In making this determination, the local processor preferably utilizes the boundary between the small stripe and large stripe regions. In an alternate embodiment of the invention, intelligence is incorporated into the device driver wherein the device driver utilizes the boundary between the small stripe and large stripe regions and incorporates this information into the WRITE command list. If an INODE
data structure is being written to the disk array 116, then in step 208 the local processor 130 builds disk specific WRITE instructions for the full stripe WRITE
operation to the small stripe region. Control then transfers to step 210, wherein the transfer controller chip (TCC) 144 generates parity data from the INODE
being written to the disk array 116. It is noted that the operation of writing the INODE to the small stripe region will be treated as a full stripe write operation, and thus no preceding READ operations associated with partial stripe write operations are encountered. Control of the operations then transfers to step 212, wherein the TCC 144 writes the data and t~e newly generated parity information to disks within the disk array 116. Control thereafter transfers to step 214, wherein the local processor 130 determines whether additional data is to be written to the disk array 116. If additional data is to be written to the disk array 116, control transfers to step 216 wherein~
the local processor 130 increments the memory addresses and decrements the number of bytes to be transferred.
Control then returns to step 206. If no additional data is to be written to the disk array 116, control transfers from step 214 to step 224 where the local processor 130 signals WRITE complete.

` 212675~ -;

If the local processor 130 receives the WRITE
command list and determines that an INODE structure is not being written to the disk array 116, then in step - ~-218 the local processor 130 builds disk specific WRITE
instructions for the data to be written to the large stripe region. It is noted that this operation requires the local processor 130 to utilize the -~-boundary between the small stripe and large stripe regions stored in the RAM 136 to develop the proper bias or offset to correct for the differing size stripes so that the proper physical disk addresses are developed. Optionally, this intelligence can be built into the device driver wherein the device driver has access to and utilizes the boundary between the small stripe and large stripe regions and incorporates this offset information into the WRITE command list. In this embodiment, the local processor 130 is not required to utilize the boundary between the small stripe and large stripe regions because this intelligence is incorporated into the device driver.
In step 220, the transfer controller chip 144 ~-generates parity information solely for the data being written. Here it is noted that if the write operation will be a full stripe write, then the disk controller 112 can generate the parity information solely from the ~-data to be written. However, if the write operation -~
will be a partial stripe write operation, then a - ;
preceding read operation may need to be performed to read the data or parity information currently on the disk. As previously discussed, these additional read operations resulting from partial stripe write operations reduce the performance of the disk system ~ ;
111. For techniques used to enhance the performance of partial stripe write operations, please see U.S. patent application serial number 752,773 titled METHOD FOR

212~7~

PERFORMING WRITE OPERATIONS IN A PARITY FAULT TOLERANT

DISK ARRAY" filed on Augu~t 30, 1991 and U.S. patent application serial number 815,118 titled'~ETHOD FOR
IMPROVING PARTIAL STRIPE WRITE PERFORMANCE IN DISX

ARRAY SUBSYSTEMS," filed ~n December 27, 199], both of which are assigned to the same assignee as this invention and are hereby incorporated by reference. In step 222, the disk controller 112 writes the data and parity information to the large stripe region. Control then transfers to step 214 where the local processor 130 determines whether additional data is to be written to the disk array 116. If in step 214 it is determined that no additional data is to be transferred, control transfers to step 224, wherein the disk array controller 112 signals WRITE complete to the disk device driver. Control then passes to step 226, wherein the device driver releases control of the system processor 20 to continue execution of the application program. This completes operation of the WRITE sequence.
Referring now to Fig. 10, a READ operation as carried out on the disk array subsystem 111 using the intelligent disk array controller 112 is shown. The READ operation begins at step 250 when the active process or application program causes the system processor 20 to generate a READ command which is passed to the disk device driver. Control transfers to step 252, wherein the disk device driver assumes control of the system processor 20 and causes the system processor 30 20 to generate a READ command list similar to that described in U.S. patent application serial no. 431,737 assigned to Compaq Computer Corporation, assignee of the present invention. The READ command list is sent to the disk subsystem 111 in step 254, after which "- ,' S

2 1 ~ ~ 7 ~

operation the device driver waits until it receives a READ complete signal.
In step 256, the disk controller 112 receives the READ command list, via l:he BMIC 142 or CPc 146 and transfer controller 144, and determines if the read operation is intended to access data in the small stripe region, i.e., an INODE, or data in the large stripe region. In making this determination, the disk controller 112 preferably compares the disk address of the requested data with the boundary between the small ~tripe and large stripe regions stored in the RAM 136 -to determine which region is being accessed.
Optionally, more intelligence can be built into the device driver such that the device driver incorporates information as to which region is being accessed in the REI,D command list. According to this embodiment, the disk controller 112 would require little extra intelligence and would merely utilize this information ~
in the READ command list in generating the disk ;
specific READ requests.
If the small stripe region is being accessed, the local processor 130 generates disk specific READ
requests for the requested INODE and its associated parity information in the small stripe region in step ;
260 and queues the requests in local RAM 136. Control transfers to step 264, wherein the requests are executed and the requested data is transferred from the disk array 112 through the transfer controller 144 and the BMIC 142 or the CPC 164 to the system memory 58 addresses indicated by the requesting task. If the disk controller 112 determines that the read operation is intended to access data in the large stripe region, the local processor 130 generates disk specific READ
requests for the requested data and its associated parity information in the large stripe region n step 2:~2~7~
, 262 and queues the requests in local RAM 136. These requests are executed and the data transferred in step 264. Upon completion of the data transfer in step 264, the disk array controller 112 signals READ complete to the disk device driver in step 266, which releases control of the system processor 20.
Therefore, by providing varying stripe sizes in a disk array, and in particular providing a region with a complete stripe size that is equivalent to the size of small data structures that are often written to the disk array, the number of partial stripe write operations are reduced. By placing these small data structures into the small stripe region where the data stripes exactly match the size of the data structure, the resulting operation is a full stripe write. This increases disk performance because the performance penalties associated with partial stripe write operations are removed.
The foregoing disclosure and description of the invention are illustrative and explanatory thereof, and various changes in the components, methods and operation as well as in the details of the illustrated logic and flowcharts may be made without departing from -~

the spirit of the invention.

;''~: ~
: ~ ~

- ~ ~
~ ~ .

Claims (11)

1. A method for managing disk array operations in a computer system disk array, utilizing parity and redundancy and recovery techniques having a plurality of data stripes of varying sizes for storing data, including a first stripe size region corresponding to a data structure type of a known size used in the disk array, the method comprising:
generating a data write operation to the disk array;
determining if the data to be written is of the data structure type;
writing the data to the first stripe size where the data is of the data structure type; and writing the data to a data stripe other than the first stripe size where the data is not of the data structure type.

2. A method according to claim 1, further comprising:
generating a read operation to request data from the disk array;
determining whether the requested data is of the data structure type;
reading the requested data from the first stripe region where the requested data is of the data structure type; and reading the requested data from other data stripes where the requested data is not of the data structure type.

3. A method according to claim 1, wherein the disk array includes a first region comprised of a plurality of data stripes having a first stripe size and a second region comprised of a plurality of data stripes having stripe size larger than the first stripe size, and the data is written to a data stripe in the second region if the data is not of the data structure type.

4. A method according to claim 3, wherein the data writs to the first region is a full stripe write operation.

5. A method according to claim 3, wherein the step of data write determining is performed by a disk controller coupled to the disk array.

6. A method according to claim 3, wherein the step of data write determining is performed by a system processor in the computer system.

7. A method according to claim 3, further comprising:
generating a read operation to request data from the disk array;
determining whether the requested data in the read operation is of the data structure type;
reading the requested data from the first stripe size region where the requested data is of the data structure type; and reading the requested data from the second stripe size region where the requested data is not of the data structure type.

8. A method according to any of the preceding claims, further comprising the initial steps of:
creating a file system on the disk array, partitioning the disk array to create the first region and the second region.

9. A computer system which performs disk array write operations, comprising:
a system bus;
a disk array, utilizing parity and redundancy and recovery techniques, coupled to the system bus having a plurality of data stripes of varying sizes for storing data, including a first stripe size corresponding to a data structure of a known size and type used in the disk array;

means coupled to the system bus for generating a data write operation to the disk array;
means coupled to the generating means and the system bus for determining if the data to be written is of the data structure type;
means coupled to the determining means and the system bus for writing the data to the first data stripe size where the data is of the data structure type; and means coupled to the determining means and the system bus for writing the data to a data stripe other than the first stripe size where the data is not of the data structure type.

10. A system according to claim 9, wherein the disk array includes a first region comprised of a plurality of data stripes having a first stripe size and a second region comprised of a plurality of data stripes having a stripe size larger than the first stripe size, wherein the first stripe size corresponds to the size of data structure type used in the disk array.

11. A system according to claim 10, wherein the data structure write operation to the first stripe size region is a full stripe write operation.