Realization of Continuously Backed-up RAMS for High-Speed Database Recovery

Transcription

1 Realization of Continuously Backed-up RAMS for High-Speed Database Recovery Yahiko KAMBAYASHI+ Hiroki TAKAKURA* +Integrated Media Environment Experimental Laboratory Faculty of Engineering Kyoto University Yoshidahonmachi Sakyou-ku Kyoto 606, JAPAN *Department of Computer Science and Communication Engineering Faculty of Engineering Kyushu University Hakozaki, Higashi-ku, Fukuoka 812 JAPAN Abstract In order to realize high performance database systems, it is very much important to develop a method for high-speed backup as well as method to realize high-speed database operations. Although semiconductor memory is getting less expensive, disks are still important to realize reliable large database systems due to their characteristics; low cost, large capacity and nonvolatileness. If system down or power failure occurs, main memory data may be lost. To avoid losing data, battery backup memories and battery backup disks are usually used. This method, however, still has problems, since each data item is not duplicated. Some system fault may change data in main memory, which cannot be recovered. If all the contents of a main memory at a certain time are stored in disks, high-speed recovery can be achieve using sequential. Continuously backed-up RAMS are designed for such a purpose. Dual-port-RAMS developed for video display systems are good candidates to realize such systems, since sequential backup can be realized simultaneously with normal read/write operations. As we can show that currently available dual-port-rams are not sufficient for our purpose, we have designed continuously backed-up RAMS using conventional RAMS. Two-plane backup RAMS and double buffer backup RAMS are designed and effectiveness of these approaches is examined by actually developing hardwired circuits. Since to realize the whole main memory by continuously backed-up RAMS is not realistic, we propose a memory system consisting of continuous backedup RAMS and conventional RAMS. High-speed recovery can be achieved by storing hotspot data (data which are modified frequently) in continuously backed-up RAMS. discussed. DATABASE SYSTEMS FOR ADVANCED APPLICATIONS 91 Ed. A. Scientific Publishing Co. 1 Introduction As demands for new application areas of database systems are increasing, development of high performance reliable database systems is very much important. Wide variety of research to achieve high performance are conducted, such as realization of high-speed operations, database machines, main memory database systems, concurrency control mechanisms and parallel/distributed database processing. This paper deals with another important problems ; high-speed recovery. For example, in transaction processing systems it is required to achieve short processing time as well as very short recovery time. A concept of continuously backed-up RAMS is introduced which can realize low overhead backup operations as well as high-speed recovery, using sequential es of disks in stead of a number of direct es. We first designed a system using dual-port-drams developed for video display systems. Contents of such RAMS can be sequentially ed during normal read/write operations. We can use a sequential port for backup operations. As discussed in Section 3.2, currently available dual-port-drams for NTSC TVs are not sufficient for our purpose. We believe, however, that even by the current VLSI technology we can develop dual-port-drams suitable for database systems. As consumer electronics have much bigger market than computer components, we can utilize dual-port-drams for the next generation high resolution TV to realize database systems economically. Discussions on such cases are also shown. In order to prove the effectiveness of continuously backedup RAMS, we have developed such a system using conventional RAMS. We designed two methods, two-pla.ne backup RAMS and double buffer backup RAMS. The former requires twice of memory elements but recovery process can be simplified. The latter corresponds to a hardware realization of the double buffer mechanism. As it requires very little additional memory elements, it is a practical method until dual-port-drams with large sequential buffer become available. Experiments shows feasibility of this approach. We need to develop recovery mechanisms for double buffer backup RAMS, since main memory contents are divided into several blocks and contents of different blocks are transmitted to disks at different time. As to realize a main memory database 236

2 system using continuously backed-up RAMS is not realistic, we also have to use conventional RAMS for main memory. In highspeed database systems it is said that a part of data (for example, 10 percent) is used very much frequently (for example, SO percent of updates are performed on such data). We think such hotspot data should be stored in continuously backed-up RAMS. Since recovery process can be realized by sequential for disks, recovery for hotspot data is realized in a short time. Recovery for non-hotspot data should be done when the data are required, so that we can restart the system right after recovery of the hotspot data only. In this way we can achieve high-speed recovery. 2 Basic Concepts In this section we will discuss necessary concepts.. Hot spot data The of database systems may concentrate in very small part of data. In typical database systems, it is known that about SO percent of concentrates in about 10 percent of database (SO-10 rule). Such data are called hotspot data.. Main memory database systems As the cost for semiconductor memory is getting lower, database systems whose main memory contains the whole database data are developed. Some critical data may be stored battery backup memory. Since to secondary storage is not required, such main memory database system attain high performance. In order to realize reliable systems, we still need disks due to their non-volatile property, since battery backup memory may have some trouble. l Commit A unit of operations to perform one job is called a transaction. By the commit operation of a transaction, data modified by the transaction are transmitted to disks. This operation may require some number of disk. In order to improve the efficiency a group commit method was introduced, where a commit operation is performed after completion of several transactions. Performance improvement is achieved by the reduction of the number of disk es, since more than one direct can be merged as one sequential.. Backup and recovery Backup operations for main memory are required to protect data from accidents which erase data in the main memory. There are the following three kinds of data required to realize recovery of main memory data after accidents. 1. Contents of main memory at a certain time (called checkpoint time). 2. Log of database operations after the latest checkpoint time. The possible latest contents of main memory should be calculated using (1) and (2) for recovery process. For large main memory we may have more than one checkpoint time [EIC88] [HAG86], which makes recovery process complicated. 3. Log disk Figure 1: The hardware structure Assumption on accidents In this paper, we consider only the case that we lose main memory data by some reason. We assume that we can recover from the database destruction of disk crush. Conventional backup mechanisms for disks are assumed to be used. Continuous Backup Using Dual- Port-RAM In this section, we discuss the organization of a backup system and its hardware structure utilizing dual-port-rams. 3.1 Organization of a backup system In order to realize a high performance system it is required to develop a backup/recovery mechanism satisfying the following requirements I. Very low overhead to get backup data 2. High speed. recovery To achieve (1) we will use dual-port-rams or memories with similar nature in order to realize backup process and read/write operations simultaneously such a memory is called continuously backed-up RAM. Sequential to disks will reduce time to use disks. We use sequential for backup process as well as recovery process. This makes recovery time very short. As the whole main memory may not be able to be realized by such continuously backed-up RAM, we will use conventional RAMS as well. In such a case we that hotspot data are stored in continuously backed-up RAMS. The hardware organization of the backup system is shown in Figure 1. The system consists of four blocks ; backup and recovery management block (BRM), database storage block (DS) in where hotspot data are stored, database management block (DM), and log management block (LM). 3.2 Dual-port-RAM To realize a backup system, dual-port-ram seems to be promising. Dual-port-RAM has two I/O ports, and its structure is shown in Figure 2. It can be regarded as a conventional RAM with another port, from which contents of RAM can be retrieved while the RAM is used. Depending on the technology used, it is classified into dual-portsram (static RAM) and dual-port- DRAM (dynamic RAM). The capacity of DRAM is bigger than that of SRAM. Basically, dual-port-dram consists of a serial port register and conventional DRAM. At a time data in 237

3 random port bit 33.3msec serial port column address (a) NTSC IMbit DRAM has four memory cells Figure 2: The structure of a d&port-dram 33.3msec 29, 6bsec scan line > 1125 line Dual-Part-DRAM V Figure 4: Examples (b) HDTV of one frame of NTSC and HDTV (a) Normal Dual-port-DRAM (b) Backup BRM Backup disk Figure 3: Dual-port-DRAM backup system one selected row are transmitted to the data register and the contents of register can be retrieved sequentially. We can use random port independently of the to serial port, except for the row whose data are under transmission to the data register. In the case of a 1 Mbit dual-port-dram, there are four 512 bits x 512 bits RAMS. Each IOW consists of 512 bits and there are 512 such rows. 4 bits can be ed parallely. In order to read all the contents in the RAM we need to the serial port for 512 times. It is enough for the application to conventional TV, but it may be too slow for database applications as discussed later. In the caee of dual-port-sram, the two random ports can be ed independently. The largest dual-port-sram available in the market currently is only 16 kbits, which is too small for database applications. In a backup system, we use a random port as an conventional main memory, and retrieve the backup data sequentially from the serial port. As normal shown in Figure 3(a), DM es a random port. During backup shown in Figure 3(b), DM es a random port, too. When BRM es serial port, one row data are transmitted from memory cells to serial port registers. After that, BRM starts to transmit the data to a disk. The advantage of this method is that DM must wait only when data are transmitted from memory cells to registers. The disadvantage is that we need a lot of checkpoints for recovery operations, i.e., the number of rows. For example, we need at least 512 data transmissions for 1 Mbit dual-port-dram. If there exist many checkpoints, BRM sends to LM many logs which have the information about backup data. This increase the overhead of log management. For example, we assume that disk data transmit speed is 800[nsec/byte], and that a main memory is organized every 4 Mbytes block (each block uses 32 dual-port-drams and lbyte consists of 2 DRAMS). We also assume that backup operations are performed in every block and that transmit operation from memory cell to register are done in all 32 DRAMS at one time. BRM send logs to LM every 6.6[msec] (= 800[nsec/byte] x 512[byte] x 16). We must take 512 logs per 6.6[msec]. Dual-port-DRAM is somewhat costly and cycle of its random port is slower than conventional DRAM, since it is developed for video memory. Current dual-port-rams cannot be used for backup systems, but dual-port-rams for high definition TV can be used for such purposes by the following reasons. In the case of an NTSC interlace scanning television system for which currently available dual-port-rams are used, one frame is constructed in 33.3[msec] (= [Hz]) and has 525 scan lines. It takes 63.5[psecj (= 33.3[msec] + 525) to scan one line. The effective scan line is 80 percent of one line and one scan line has 512 pixels (serial port register has 512 bits). The read cycle of a memory is 63.5[psec] x 0.8 e-512 = loo[nsec]. We must use RAM whose read cycle is under loo[nsec]. On the other hand, read cycle of one serial port is 40 to 6O[nsec/bit]. Therefore, we can use dual-port-dram for video memory without conventional method such as shift registers for a frame memory which utilizes conventional RAMS. Although dual-port-dram is useful for video memory, we cannot use our backup system since the capacity of its serial port register is too small. In case of HDTV(High Definition Television), we assume that one frame is constructed in 33.3[msec] and has 1125 scan lines, and that one scan line has 1333 pixels (total n,umber of pixels for one frame is : 1333 x 1125) by Japanese NHK standard: We must use RAM whose read cycle of serial port ( is under 17.8[nsec]. This means that dual-port-dram needs to have faster serial port. Although read cycle of serial port becomes faster, disk data transmission speed is not 238

4 Backup-d<sk (4 N ormal (a) Normal DM L--J Backup disk (b) --._ DM (b) transmit - =r 4iffer memory plane R / I Backup disk (c) transmit BackuT-disk Figure 5: Two-plane backup system fast enough. To deal with such large data quickly without conventional method, dual-port-dram needs faster random port and larger serial port register : e.g bits to 1333 pixels. When the capacity of serial port register became larger, we can reduce the number of logs which have the information about backup data. If dual-port-dram is developed for such a purp9e, we can organize our backup system easily. We believe that even by the current technology, dual-port-ram suitable for backup memory can be designed. We should, however, use mass produced products in order to reduce costs. 4 RAM Based Continuously Backed-up Memory As discussed in the previous section, current dual-port-rams are not suitable for our purpose, we will design a memory system with a similar function using conventional RAMS 4.1 Two-plane backup The system, shown in Figure 5, uses memory plane (MP) and backup memory plane (BUMP). In normal shown in Figure 5(a), DM es both MP and BUMP. Both MP and BUMP store identical data. During backup shown in Figure 5(b), DM remains to MP and BRM es BUMP in order to back up the BUMP data to the backup disk. Although this backup requires a little time, the data of MP and these of BUMP may become different after the backup. Thus, we must transmit the data from MP to BUMP, as shown in Figure 5(c) in order to make the both contents identical. At the same time, DM also es both MP and BUMP, since database processing (c) Backup Figure 6: Double buffer backup system continues. The data transmit between MP and BUMP influences the speed of DM. The advantage of this system:is that only one checkpoint time exists so that we can recover from accidents easily. The disadvantage is, powever, that this system needs twice of memories, compared with a conventional system. 4.2 Double buffer backup Since the two-plane backup system requires too much memory, we have designed a double buffer backup system which simulate a dual-port-dram. As shown in Figure 6, it corresponds io a hardware realization of a double buffer method. In our system, a memory plane consists of 1 Mbyte DRAM module which consists of eight I-Mbit-DRAMS, where one module has two buffers, and each buffer consists of 32-kbytes-SRAMs. The double buffers are controlled by hardware. As normal shown in Figure 6(a), DM es a memory plane. Before backup starts, a part of data are transferred from memory plane to one buffer, as shown in Figure 6 (b). After data transmission, BRM es the buffer in order to back up the buffer data to backup disk. At the same time, another part of data are transferred from a memory plane to the other buffer, as shown Figure 6(c). These operations are repeated until the backup finishes. The advantage of this system is that the cost is the minimum among the three, since the capacity of this system is times as large as that of conventional system. There exist 32 checkpoints.which are much fewer than dual-port-dram. As discussed in Section 3.2, BRM send logs to LM every 105[msec] (= 800[nsec/byte] x 32[kbyte] x 4). We must take 32 logs per 105[msec]. In this system, most of memory data ate transferred while data are backed up. The disadvantage is that the control is somewhat difficult, because the backup data cannot transmit from memory plane to buffer at one time as dual-port-dram does _.- --_ ^....r-& r-j, _^ _-._ -.

5 Figure 8: Main memory organization (a) Normal (b) Backup Figure 7(b), we select a request from DM as a trigger, and then we cannot see any signals of backup because of asynchronous. In Figure 7(c), we can see all signals. Because DM and transmit are done synchronously. If both DM and BRM request at same time, either DM or BRM is waited until the other request ends. We estimate the working time of our system by these results. In this paper, we assume that our system has 16Mbytes continuously backed-up main memory which are divided into 4 Mbytes blocks, that the time of data transmission from memory to disk is SOO[nsec], and that disk response time is 20[msec]. Every 4 Mbytes part of the main memory works parallely. The total time to back up 4Mbytes data is 4[Mbyte] x SOO[nsec/byte] + 2O[msec] N 3.4[sec] During transmission, we assume that the data are transmitted by 2 bytes units. If DM es a main memory uniformly, the average transmit cycle is 650[nsec]l. The time required to transmit 4 Mbytes data is i[mbyte] x 650[nsec/byte] N 1.4[sec] (c) transmit Figure 7: The results of testing a prototype circuit 4.3 Experimental results We have developed a prototype system using conventional RAMS which is introduced in subsection 4.1. The capacity of memory plane is 128 kbytes. The results are shown in Figure 7 and Table 1. In Figure 7(a), a request of is sended from DM to DS and then an acknowledgment of this request is sended from DS to DM. This result shows normal. In Figure 7(b), a request of is sended from DM to DS and then an acknowledgment of this request is sended from DS to DM. This is the same as Figure 7(a). It is, though, different that a request of is sended from BRM to DS and an acknowledgment is sended from DS to BRM. DM and BRM are sended asynchronously and acknowledgments of these are also sended asynchronously. In mode normal transmit backup time [nsec] cvcle lnsecl Table 1: Time required to perform operations in the prototype system This data transmit time is smaller than backup time. Thus if we use double buffer backup, we need not consider the data transmit time. We discuss the comparison of our backup system with other conventional backup systems in section 6. 5 Organization of Database Systems Utilizing Continuously Backed-up RAMS 5.1 Main memory organization In conventional main memory database systems, updated data are written in database disks, not distinguishing hotspot data from non-hotspot data, in order to recover from accidents. To overcome disk delay, many systems uses buffers, disk array etc. These devices are efficient if write rate is not so high. There are, however, database machines of high write rate. In transaction processing systems, the requests of write occur frequently. If we use such high write rate systems, the buffer may frequently overflow and cannot recover from overflow in a short time. Read must be delayed for long time until at least one write ends. Because disk response time of random is far slower than CPU speed : e.g., about a hundred The probability that one memory block is ed is about 0.25.Thus average time is 950[nsec] x [nsec] x (l-0.25) = 650[nsec] 240

6 hatspat data non hotspot data accident I +-time happen Figure 9: Different status for hotspot data and non-hotsp data ot j %, 27FFFF, time Figure 10: Backward search thousand times slow, such systems cannot utilize the efficiency number of 104s T ia of main memory database enough and may be not different from disk-based database. If some parts of main memory are realized by battery backup RAM and/or EEPROM (electronically erasable programmable 1 / [B, 27FFFF, U- no u.pdate ROM), we can write updated data to such memory part to im- At ime prove reliability. Although battery backup RAM is very efficient and easy to realize, it cannot be so safe as disks since the data Figure 11: Recovery operations stored in RAM may be modified by accidents OP due to the life of the batteries. Because the time of EEPROM is much [B, O IFFFF, means that main memory data from slower than RAM, EEPROM reduces efficiency of a main mem- to OPFFFF are backed up at this time and last transaction ID ory. is The type 3 log shows that transaction Tid begins at We need the database systems which are not influenced by this time. Finally, the type 4 log shows that transaction Tid disk, battery backup, or EEPROM. As discussed in Section 3.1, terminates at this time. we distinguish hotspot data from non-hotspot data. Hotspot data cannot write in database disk because random of hotspot data cause a buffer overflow even if very large buffer is used. 5.3 Recovery operation We write all hotspot data from top to bottom address in backup In conventional systems, recover from accidents takes long time, disks which are used to store hotspot data exclusively. Because of because all operations which were not reflected in disks are resequentially write, we can neglect the response time, as discussed done immediately after accidents and these operations need many in Section 4.3. random disk es in order to keep consistency of all data. As The main memory organization is shown in Figure 8, where an alternative to this, we do recovery operations for hotspot data non-hotspot data or temporary data are stored in regular main immediately by sequential, and for non-hotspot data at the memory and are transmitted to database disks after commit. time when these operations are needed. The operation to recover Hotspot data are stored in main memory with continuously hotspot data requires time much less than that of conventional backed up system, DS, which is discussed in Sections 3 and 4. system. 5.2 Type of log information Recovery process for hotspot data is as follows. At first, we transmit the backup data from the backup disk to DS. At the same time, we write the log which have been left in the buffers In our system, hotspot data are written in backup disk periodisince accident happened into log disk. Next, we search the logs cally and non-hotspot data are written in database disk at every backward and check type 4 log, [Z id, E], to set a bit of a bitmap commit operation of a transaction. Since the time when data are backed up is different for each kind of data as shown in Figure 9, the recovery operations for both data are not simple. We need to get the hotspot data states just before accident happened. In our system, we use the log of the following four types. l type 1 log : [ I id, M, Ad, Vail, Va/2] l type 2 log : [B, Ad,Tid] 0 type 3 log : [Tid,S] l type 4 log 1 [Tid,E] The type 1 log shows that transaction whose ID is I id updates the data of address Ad from Vail to Va12 and M is the byte length of updated data. Transaction ID is given in order. The type 2 log shows that data of address Ad mark begin to he backed up at this time and the last transaction ID is Tid. For example, on until the type 2 log, [B,Ad], appears twice. For example, as shown in Figure 10, we begin reading logs backward. Type 2 log, [B, 27FFFF, OOA9], appears at first. We continue to read, then logs and type 2 log, [B, IFFFFF,0067], [B, I7FFFF,OO5B],... appear one after another. When type 4 log [OOAl,E] appears, we set a bit of the transaction OOAI in bitmap 1. At last, we find type 2 log, [B,27FFFF,0007], again. When the identical type 2 log appears, we can read logs forward. Transactions which began before the checkpoint and ended between the checkpoint and accident time may exist. In such a case, we check type 2 log [B, 17FFFF,0007]. We read bitmap between transaction ID=0007 and the last ID at the time when accident happened. When we find out that T&$ s bit is 0, we must undo operations which this transaction has done before the checkpoint [B, 27FFFF, 0007) Then we can begin to read logs forward. As shown in Figure 11, when the type 1 log, [T& M, Ad, Vail, VU/~], appears, we read Tid of type 1 log and check the bitmap. If the 241

7 bit fo; Tid is l, the data is updated. If the bit is O, the data isn t updated. These operations are done until the logs run out. These operations do not recov& the data which were updated by non-terminated tra&.actions. If recovery of such data is recjuired, we need not check a bitmap and would recover with all type 1 logs. 6 Comparison with Other Methods In this section, we discuss the efficiency of our backup system comparing with other methods. 6.1 Comparison with the shadow paging method Using similar analysis which we discussed in subsection 4.3, analysis of random backup is as follows. We assume that one page consists of 1 kbyte and backup is random write to database disk. This system takes (l[kbyte] x 80O[nsec/byte] -t 20[msec]) x lb[mbyte] 1 [kbyte] cz 341[sec] The result shows that our backup system is more effective than random backup, as far as backup time is concerned. To avoid random, shadow paging [SON871 is presented. In shadow paging, backup data are written in database disks behind usual read or write. This strategy makes scheduler overhead become larger unless effective scheduler is introduced, and makes throughput of a database system become worse. Our backup system writes backup data in backup disks using hardware controlled scheduler (logic circuit). This scheduler little affects usual scheduler which controls normal read or write to database disks. 6.2 Comparison with the safe RAM method The backup systems introduced here can be used as Safe RAM [COP89]. Safe RAM consists of memory and disk with battery backup. Reliable writes are written in Safe RAM and are stored from Safe RAM to disk with background operations. When accidents happen, reliable writes begin to store from Safe RAM to battery operated disk. Safe RAM is classified into the Separate Safe and Integrated Safe. Separate Safe consists of a main memory, RAM which is separate from a main memory, and disk. RAM and disk are backed up by battery. Integrated Safe consists of a main memory, which is treated as Safe, and disk. All memories and disk are backed up by battery. Although our backup systems are similar to Separate Safe, efficiency of recovery is very much different. Both write updated data into a disk in background. Safe RAM uses a database disk, which is also ed by main memory, in order to reflect reliable writes into a disk and Safe RAM requires random. Since disk determines the processing time, the response time can be very long. In contrast to this, our system uses a backup disk which is only for backup and is ed sequentially. Because of the sequential, our system needs less buffer size than Safe RAM (in the case of the double buffer backup) and takes less time in order to write data into backup disk. At recovery, Safe RAM takes more time than our sequential backup systems. Because Safe RAM must recover main memory data from database disk with random. But our systems read from backup disk sequentially for hotspot data. 7 Concluding Remarks In this paper, we have shown backup systems for main memory database systems. Since such a system is still expensive, we recommend to use it for frequently updated data, i.e. hotspot data. We discuss three types of backup systems ; using dualport-dram, two-plane backup, and double buffer backup. These systems execute sequential backup of main memory data into a backup disk. Current dual-port-dram is not useful, due to the cost, time, and the size of serial port registers. If more useful dual-port-drams such as these for HDTV are developed, they can be utilized. Recovery operations against accidents are also discussed. These operation are shown to be more effective than those for conventional systems for hotspot data. The case when hotspot data change due to the usage change is not considered in this paper. Acknowledgment The authors thank Mr. Keizo Saisho at Kyushu University and Mr. Chiaki Nakamura at Nagasaki University for helpful discussions. The authors thank Mr. Martin Santavy for correcting English of this paper. This work is supported in part by the Science Foundation Grant of the Ministry of Education and Culture in Japan. References [COP891 [EIC88] [HAG861 [SON871 G.Copeland, T.Keller, R.Krishnamurthy, M.Smith, The Case For Safe RAM, Proceedings of the Fifteenth International Conference on Very Large Data Base, Amsterdam, The Netherlands, August, 1989, pp M.H. Eich, MARS : The Design of a Main Memory Database Machine, Database Machines aud Knowledge Base Machines, Kluwet Academic Publishers, 1988, pp R.B.Hagmann, A Crash Recovery Scheme for a Memory-Resident Database System, IEEE Trans. on Computers, Vol. C-35, No.9, September, 1986, pp S.H.Son, A Recovery Scheme for Database Systems with Large Main Memory, COMPSAC, Tokyo, Japan, 1987 pp