Performance Analysis of RAID in Different Workload

Transcription

1 Send Orders for Reprints to 324 The Open Cybernetics & Systeics Journal, 2015, 9, Perforance Analysis of RAID in Different Workload Open Access Zhang Dule *, Ji Xiaoyun, He Miao and Zhu Huaijie China Petroleu Pipleline Engineering Corporation,,,, P.R. China Abstract: A perforance evaluation odel is built for the RAID syste with queuing network. With MVA ethod we develop, validate and apply an analytic perforance odel for disks arrays configured as a RAID 5. The results show that the odel can basically express the trend of the perforance with different load. Keywords: Analytical perforance odel, array cache, disk array, I/O perforance evaluation, parallel I/O, RAID. 1. INTRODUCTION The disk array which has the characteristics of large storage capacity and high perforance is the typical secondary storage. With the increasing iportance of disk arrays, odels for evaluating their perforance have becoe increasingly iportant. Unfortunately, the actual syste is very coplex, the perforance is affected by any factors such as hard disk perforance, RAID level, different load types, RAID controller etc. Thus, defining an analytic perforance odel proved far ore difficult than expected. In this paper, we develop, validate and apply an analytic perforance odel for disks arrays configured as a RAID 5. In order to the siplicity and coputational efficiency, the perforance odel uses the average value for fork-join networks analysis technology. Another advantage is that the predictions of odel are right verified in a real syste, the HP Sart Array P410i. As verified in a real array, the predicted values under the tested workloads are, on average, accurate within 3.32% of the real values. Hence, our odel is siple and intuitive. 2. RELATED WORK Response tie and throughput are the key perforance etrics of disk array. Response tie is the total aount of tie taken to respond by an I/O request at the disk array. Throughput is the rate at which the disk array can service Input/Output operations, it is usually easured in Input/Output operations per second (IOPS). Research on the odel of RAID perforance can be traced back to last century 70's. Siulation techniques and analytic techniques are the ost coon ethod built Perforance odels. Papers [1] ainly suarizes four priary categories: (1) siulation ethod; (2) analytic odels that deal with fork-join networks and neglect queuing effects; (3) analytic odels that deal with fork-join networks and neglect its synchronization; and (4) restricted queuing E-ail: cppe_zhangdl@cnpc.co.cn odels that the odeling technique is based on fork-join synchronization using the MVA techniques. The following lists analytical odels of disk arrays that are published in recent years. Paper [2] describes an object RAID syste by using Object-based storage Devices and presents queuing odels with cache to analyze the syste; Paper [3] presents a odel of I/O service tie and the best perforance; Paper [4] presents an analytical odel that incorporates the factors of the real syste, such as array caches, array controllers, disk controllers, and so on; Paper [5] presents a perforance evaluation odel that is built for the RAID software syste with queuing network. Hence, the perforance odels for disk arrays in the papers are alost very coplicated. In this paper, we use the top-down design ethod to odeling RAID syste. We analyze and study the principle and process of the RAID syste, and identify the key coponents, and explicitly odeling the. The perforance odel will allow us to copute the entire syste's throughput and response tie by using the MVA technique. 3. PERFORMANCE MODEL A real disk array is coposed of hardware and operating syste, the hardware of the disk array is ade up of several parts, for exaple disk caches with their own internal caches, array caches, host interfaces, array controllers, disk interface, disk (SAS II) controllers, internal buses, etc. So, it is a very coplex. In order to develop an analytic perforance odel, let us identify the key coponents of the syste. Let us analyze and study how the disk array services its workload. All I/O requests fro the client are first subitted to the array controller. A read request is straight away returned fro the array cache if the array cache is hit in; otherwise it is subitted to the disks. A write request is first written into the array cache and later subitted to the disks. Hence, All requests are first processed by the array controllers and subitted to the array cache and array disks. Based on this analysis, three coponents including the array controller, the array cache and the array disks are identified as the key coponents, so the odel is ade up three nodes X/ Bentha Open

2 Perforance Analysis of RAID in Different Workload The Open Cybernetics & Systeics Journal, 2015, Volue Fig. (1). Queuing network odel of disk arrays with cache. Fig. (2). Overlap of bus phases and echanis. We abstracts the array cache into a single queuing server, and abstracts the array controller into a fork-join queue, and abstracts each disk into a queuing server. Fig. ( shows a fork-join network odel of disk arrays with cache that odels with M jobs that cycle between the M terinals and the disk-array CPU Service Tie Based on equipent using single CPU and pressure test, CPU node is abstracted into FCFS queuing node. The CPU service tie for all tasks is a constant and is given by T CPU_Delay = constant Cache Service Tie In a real syste, cache is a saller, faster eory that stores data for iproving overall syste perforance, the access tie that an I/O request can be served by the array cache is faster than served directly by a set of disk drivers. Hence, the ore I/O requests hit the array cache, the faster the disk array perfors. The Cache service tie is the total aount of tie which the data are transferred between the array cache and the client, it is given by [6]: ST buf (b) = b V RAM 3.3. Disks Service Tie The disks are a data storage device and are one of the ost iportant parts in the disk array, it has the characteristics of slow speed and high reliability, it is used for storing and retrieving data. hus, the disk service tie is the ain constitution of the response tie, it is necessary to get the predictions of disk service tie. Paper [7] falls into the SAS disk drive odel and working principle, it is given by Fig. (. The disks service tie (T s ) is the su of disk positioning tie(t p ) [4], transfer tie(t t ), and parallelis overhead for N disks (To) [8]. T s =T p T t T o Where, the Variable T s, T p, T t, T o is: T t (b) = b V disk

3 326 The Open Cybernetics & Systeics Journal, 2015, Volue 9 Dule et al. T p = a b disk _ queue > 3 1 disk _ queue c d * disk _ queue disk _ queue 3 where the constants are a =3.53 s, b =8.81 s, c =-2.73s and d =3.68s. T O = (0.5 1 n 1 ) S 3.4. Perforance Model According to the analysis result, the perforance odel used by the MVA technique is expressed as: R()=T wait ()T service () T wait ()=R(-1) R(0)=0 R()=*T service (1)(-1)* T service (2) T service () = ( i 1) T service ( i) i= 1 T_service()=T CPU ()T cache ()P()*T disk () Where the variable eans nuber of I/O workload streas, it varies fro 1 to M. The R() eans response tie that I/O request is serviced by the disk array. The P() eans the probability that the request can be served directly by a set of disk drivers, it depends on the cache size of array disk, and on the I/O workload and so on, the predictions of P() is very difficult. P() = 1-hit _ probability = 1 served fro a disk directly 0 otherwise The T disk () eans the disks service tie, under the sae workload, the sae the disk services is, naely, T disk ()=T disk (-1)=T disk, the response tie of the request is given by R() = ( P i 1) ( 1) T 2 CPU P i ( 1) B Size 2 V Cache T Disk 2 The response tie of the disk arrays R is given by: R = R() = 1 2 T CPU 1 2 B Size V Cache ( P i 1) 2 P i T Disk 3.5. Read Model The HP Sart Array P410i uses a read-ahead technology to iprove overall syste perforance by predicting when sequential read requests will follow [9]. With the cache hitting, the read requests is served by the cache, the data is returned fro the array cache and the disk array control signals service copletion. Otherwise, the request is served by a set of disks and all disks I/Os for the request finish task, the request is finish Write Model The HP Sart Array P410i uses a write-back caching technology (FBWC) to iprove overall syste perforance Fig. (3). FBWC block diagra.

4 Perforance Analysis of RAID in Different Workload The Open Cybernetics & Systeics Journal, 2015, Volue by writing directly to the cache, this writing approaches is ore coplex to ipleent. With technique of write-back caching, write coalescing, full-stripe writes, and readodify-write, The HP Sart Array P410i significantly iproves perforance for I/O operations. FBWC architecture is shown in Fig. ( [9]. operating syste, is used to generated the workloads and access the HP Sart Array P410i. A HP Sart Array P410i array with 3.5inch 300GB SASII disks, one controller, and 256 MB of cache. HP Sart Array P410i containing 5 disks. The LUN is configured at RAID 5 and uses a stripe unit size of 256 KB. Fig (4). Predicted and Measured Response tie for Job=1. Fig (7). Predicted and Measured Response tie for Job=4. Fig (5). Predicted and Measured Response tie for Job=2. Fig (8). Predicted and Measured Response tie for Job=5. Fig (6). Predicted and Measured Response tie for Job=3. 4. EMPIRICAL VALIDATION To validate the analytic odel developed in the previous section, we copare the predictions of analytic perforance odel against the siulation results over the range of syste and workload paraeters investigated. The disk odel is based upon the HP Sart Array P410i. We use HP DL380G5 server as a client. A HP DL380G5 server with four 3.2GHz Intel E4250 processor and 16GB ain eory, running the Microsoft Windows Server 2008 Fig (9). Predicted and Measured Response tie for Job=6. We use synthetic workloads with I/O request sizes ranging fro 4 KB to 2048 KB. The nuber of jobs generating requests range fro 1 to 8. Ioeter is selected for the siulation tool. Ioeter is a storage subsyste benchark and troubleshooting tool. Ioeter is one of the ost coonly tools. It can show the CPU utilization rate, the axiu throughput, the response of the disk syste. In order to siulate the actual environent, Ioeter allows users set different paraeters, for exaple, the access type (such as sequential, ran-

5 328 The Open Cybernetics & Systeics Journal, 2015, Volue 9 Dule et al. CONCLUSION The paper presents an analytic perforance odel that deal with fork-join networks and incorporates the effects of a real array syste. We first identify and odel the key coponents, then odel the disk array as fork-join network, derive the response tie and throughput calculation forula by using the parallel MVA technique. The validity of the odel with a specific disk array(the HP Sart P410i) is proved for a variety of synthetic workloads, request size range fro 4KiB to 2048KiB, the nuber of job range fro 1 to 8. Through estiation, the predictions of odel indicate that the real throughput values for P410i under the rando and sequential workloads is quite siilar. Across all workloads and experient sets, the predicted values are, on average, within 3.32% of the real values. However, the difference between actual and predicted values under the specific workloads(request size=sus and nuber of jobs >2) is uch higher (24.23%), These error argins are reasonable since the details of stripe unit and optiizations are not revealed by the anufacturers. CONFLICT OF INTEREST The authors confir that this article content has no conflict of interest. Fig (10). Predicted and Measured Response tie for Job=8. do), the block size, the queue depth and so on. It has the following characteristics: siple operation, recording the test script, the graphical user interface. Thus, the software can accurately reflect effectively the perforance of the storage syste. Figs. (4-10) present the redicted response tie and the easured response tie for read and write workloads for a range of request sizes, ultiprograing levels, think ties, and sequentially degrees. ACKNOWLEDGEMENTS Declared none. REFERENCES [1] E.K. Lee, and R.H. Katz, An Analytic Perforance Model of Disk Arrays and its Applications, Tech. Rep. UCB/CSD91/660, University of California at Berkeley, [2] G. Liu, J. Zhou, M. Jiang, and K. Wang, Perforance analysis of object RAID, Coputer Science, vol. 32, no. 12, pp , [3] Q. Chen, and J. Zhang, Analyse of I/O Service tie in high perforance disk array, Mini-icro Systes, vol. 21, no. 3, pp , [4] E. Varki, A. Merchant, J. Xu, and X. Qiu, Issues and challenges in the perforance analysis of real disk arrays, IEEE Transactions on Parallel and Distributed Systes, vol. 15, no. 6, pp , [5] Y. Zhang, Y. Hu, and T. Jiang, Perforance evaluation odel for RAID storage systes based on queuing Network, Journal of Changchun University of Technology (Natural Science Edition), vol. 31, no. 4, pp , [6] G. Xie, J. Liu, G. Wang, X. Liu, and J. Liu, Perforance evaluation odel for Exp-RAID syste based on MCQN, Journal of Coputer Research and Developent, vol. 45, no. z1, pp , [7] C. Rueler, and J. Wilkes, An introduction to disk drive odeling, Coputer, vol. 27, no. 3, pp , [8] K. Zhou, J. Zhang, and D. Feng, Analyzing of I/O response Tte and throughout of RAID with cache, Microelectronics & Coputer, vol. 20, no. 8, pp , [9] HP Sart Array controller technology [EB/OL], baidu.co/link?url=dkzizwaibszwx_qtn7yb9kjcaq5v1jlru- HW58CNGuTyPZ4xQl0ZS74f6iwdSDS4XHHgIj6casFb2tiGeAH_ vxtxdlto1urdj4j1okltbo Received: Septeber 16, 2014 Revised: Deceber 23, 2014 Accepted: Deceber 31, 2014 Dule et al.; Licensee Bentha Open. This is an open access article licensed under the ters of the Creative Coons Attribution Non-Coercial License ( licenses/by-nc/3.0/) which perits unrestricted, non-coercial use, distribution and reproduction in any ediu, provided the work is properly cited.