Architecture of storage area network system 

Context in source publication

Context 1

... out SCSI commands and messages from SCSI command queue and message queue respectively. If SCSI command instructs to read or write large data, the processing thread of SCSI target simulator then passes the command to the SCSI sub-system of the immediate lower layer, which fully processes the command, performs SCSI disk operations and waits for the sub-system to return results. As to other commands, the procession thread of SCSI target simulator transfers the command to I/O transferring layer, where the command is translated to block I/O requests and passed to RAMDISK driver for memory procession. The devices in SCSI target simulator are mapped to SCSI disks (Target: Lun) in initiator one by one. Here one single RAMDISK device is adopted as a whole instead of allocating each SCSI disk device to a certain RAMDISK device. Therefore, RAMDISK is divided into many different sections, and there can be different ways to map RAMDISK section to SCSI disk device, either one of the former to all of the latter or multiple of the former to all of the latter, namely 1:N, M:N, N:N. Here N:N reflection is applied, in which each RAMDISK section is mapped to one small file on the SCSI disk or file property operation. The asynchronous way applied here to process the SCSI command by the thread greatly promotes efficiency. The concurrent modification of RAMDISK property in the device-mapping table of the SCSI target simulator makes reallocation of the RAMDISK device and section convenient. On the other hand, since the speed of operations on RAMDISK is very high, synchronous mode can be used to store and fetch data. Thus the sequential operations of the RAMDISK at the client are ensured, as well as the consistency of the RAMDISK with the file system is guaranteed. In the implementation, the critical sizes for each data block to read and write operations can be flexibly and separately decided corresponding to the statistics attributions of the file system workload in application. Generally the size of RAMDISK is constant and is divided into multiple sections, each corresponding to a small file of SCSI disk or file attribution operation. If the threshold value is set too large, then the RAMDISK area in each SCSI disk will be consumed too quickly. Correspondingly, if it is set too small, most reading and writing requests will then be sent to SCSI disk, affecting the system efficiency as a whole. For an instance, for database and web server application, the size of blocks is generally 8KB, take up to 50% of the whole room, the threshold value can be set to 16KB and RAMDISK can be divided into 8 parts, each with a capacity of 64MB, thus the system gains a quite good efficiency. On the other hand, a larger threshold value can be set when the much more large files are requested, for example, in VOD service system. In future system, not only can RAMDISK be of large capacity, but dynamically auto-adjust threshold method can be applied as well. Since RAM, a volatile media, is one of the main storage devices in such an FC-SAN system based on physical disks and RAMDISK, special process should perform on the RAM-based DISK under circumstances of abnormality, normal logoff and initialization. Therefore, in the real system, a DOM with a capacity of 512M is selected as the perduring storage media for the RAM-based DISK, thus effectively files copy in response to continual I/O requests. At initialization of the system, program dd (a block copy program under the UNIX- like system) is applied to completely contents copy of DOM directly to the RAM- based DISK. When the system is running, three methods can be adopted for the files copy, file compression method, directly complete copy method and LOG file system. In the real system, the second method above is applied, with a speed of one copy per 60*6 seconds, namely a capacity of 512M from RAM to DOM per 60*6 seconds. A copying thread running in background manipulates all of the tasks of data copying from RAM to DOM. In addition, DOM is an IDE DISK based on FLASH, so the copying task of 512M from RAM by using the background copying thread occupies very little CPU resources and therefore has no impact on the system efficiency. In actual application, there is tradeoff between copy window and process ability because data inconsistency may occur when interval between two consecutive tasks of copying is large. The testing system is made up in laboratory environment. 7 Xeon nodes are used as server nodes which are the clients of the SAN visiting the storage resources. Each node has 4CPUs, SMP mode, 1G DDR RAM and 36G Disk. The operating system is RedHat Linux 7.3. FC network with 2Gbps bandwidth is adopted as resource data transfer network. There is one storage node in the system, which has 2 Xeon 2.4G CPUs, 1G memory and 7*73G disk array. Single SCSI bus connects the disks and the storage node. Special embedded operating system is used in the storage node. The testing hardware environment is shown in Fig 4. The 7 physical disks in disk array are mapped to front host nodes respectively through storage network, that is to say, every host has one network disk to operate. This mapping procedure is transparent to the host file system, and the network disk operates the same with local disk logically. IOMETER is used as testing software. There are two testing group, one tests the system without RAMDISK technique; the other tests the system with the RAMDISK technique. The threshold size of files is set to 64KB. We suppose that RAMDISK has been filled with the files to read before the testing starts. All the file requests to the data smaller than 64KB are done by RAMDISK. The results are shown in Fig.5. We can see from the results that the smaller the files are, the worse the system performance is without RAMDISK. The performance is increased much after adopting RAMDISK technique. It is because that, large quantity of small file operations consumes quite much time on seeking. Especially when the files are very small and distributed discretely, there is bigger expense in disk read performance. The size of files is still set to 64KB, and sequential mode of writing is used. The results are shown in Fig.6. The results show that, writing performance is obviously increased after using RAMDISK. The increasing is similar in the area of 2KB to 64KB, but not very obviously. The reason is that, the RAID subsystem in storage node use cache write- back technique. By such technique, an I/O operation is completed when the data have been written into the memory of SCSI-RAID adapter, and the firmware of SCSI- RAID adapter will move the data to physical disk later to get better write performance. Generally, the size of SCSI-RAID memory assigned to one disk is 64KB, so write operations about the data smaller than 64KB can be done in the cache. That is why the write performance is better than read without RAMDISK. On the other hand, 200MB/s bandwidth of FC also limits the improvement of write performance. With the RAMDISK, the performance is improved 10%~40%, and reaches 185MB/s, 92% of FC bandwidth. It is near the theoretic value. An improved FC-SAN based on the rules of file system workload is designed and implemented which uses both RAMDISK devices and DISK devices as the storage devices in network storage system. By evaluating the load heavy or light, the system selects DISK or RAM devices as the main media for I/O operation, file operations of larger data will be done on disk devices, and reading/writing smaller data or file property operations will be done in RAMDISK devices. The evaluation results reveal that, without RAMDISK, the smaller the file operation is, the worse the performance is. To the system above, the performance of read is 124MB/s and the one of write is 164MB/s. after using RAMDISK, the performance of read is improved to 170MB/s and the one of write is improved to 188MB/s. So, we can conclude that: After using RAMDISK, read performance is improved 50% to 100%, and write performance is improved 10% to 40%. Write operations can make use of 92% bandwidth of Fibre Channel. The more read requests in mixed requests and the smaller the blocks are, the more the performance of the system can be improved. This system reduces the time in seeking in disk and operation times of physical disks, decreases the speed gap between CPU and disks and improves the performance of the whole system. Acknowledgement. The works described in this paper are supported by National High-Tech Research and Development Plan of China under Grant ...