Ths paper appeared n the Proceedngs of the 2th IEEE Symposum on Computers and Communcatons (ISCC'08, Marrakech, Morocco, July 2008. Improvng Relablty and Energy Effcency of Dsk Systems va Utlzaton Control Kranma Bellam, Adam Manzanares, Xaojun Ruan, Xao Qn *, and Ymng Yang Department of Computer Scence and Software Engneerng Auburn Unversty Auburn, Alabama 36830, USA {kzb0008, acm0008, xzr000}@eng.auburn.edu, xqn@auburn.edu Intel Corporaton Ro Rancho, NM 8724 ymng.yang@ntel.com Abstract As dsk drves become ncreasngly sophstcated and processng power ncreases, one of the most crtcal ssues of desgnng modern dsk systems s data relablty. Although numerous energy savng technques are avalable for dsk systems, most of energy conservaton technques are not effectve n relablty crtcal envronments due to ther lmtaton of gnorng the relablty ssue. A wde range of factors affect the relablty of dsk systems; the most mportant factors dsk utlzaton and ages are the focus of ths study. We buld a model to quantfy the relatonshp among the dsk age, utlzaton, and falure probabltes. Observng that the relablty of a dsk heavly reles on both dsk utlzaton and age, we propose a novel concept of safe utlzaton zone, where energy of the dsk can be conserved wthout degradng relablty. We nvestgate an approach to mprovng both relablty and energy effcency of dsk systems va utlzaton control, where dsk drves are operated n safe utlzaton zones to mnmze the probablty of dsk falure. In ths study, we ntegrate an exstng energy consumpton technque that operates the dsks at dfferent power modes wth our proposed relablty approach. Expermental results show that our approach can sgnfcantly mprove relable whle achevng hgh energy effcency for dsk systems.. Introducton Dsk drves started the new era n the computer dctonary. Therefore, t s crucal to have hghly relable, energy effcent and cost effectve dsks. Extensve research has been done (and contnues to be done to reduce the energy consumpton of dsks. Lterature proves that very lttle or no research has been done to maxmze relablty and energy effcency of the dsks. Relablty may be the most mportant characterstc of the dsk drves f the data stored on the dsk drves s msson crtcal. Dsk drves are generally very relable but may fal. In ths paper dsk age and utlzaton are studed as the major factors that affect dsk drve relablty. In ths paper we studed the falure probabltes wth respect to dsk age and utlzaton. Ths relatonshp s made explct n the form of graphs. These graphs are used to estmate safe utlzaton levels for dsks of dfferng ages. Safe utlzaton zones are the range of utlzaton levels where the probabltes of dsk falures are mnmal. Smulatons prove that when dsks are operated n the safe utlzaton zones, the probabltes of falures are effectvely mnmzed. It s evdent that dsk drves consume large amounts of energy. Reducng the energy consumpton not only lowers electrcty blls but also tremendously reduces the emssons of ar pollutants. Energy consumpton s reduced by operatng the dsks at three power modes: Actve, Idle and Sleep. Mrrorng dsks (a.k.a., RAID s consdered for the experments, whch uses a mnmum of two dsks; one prmary and one back up. The data s mrrored to the backup dsk from the prmary dsk. Tradtonal methods wake up the backup dsk when the utlzaton of the prmary dsk exceeds 00 percent. Load balancng technque keeps both the prmary and back up dsks always actve to share the load. These methods consume a massve amount of energy, as the dsks stay actve even when there are no
requests to serve for a long perod of tme. The relablty of these dsks s also gnored most of the tme. We desgned a polcy where we operate the dsks at dfferent power modes based on the utlzaton of the dsk. The utlzaton of the dsk s calculated a pror. Along wth the power conservaton we also am at achevng hgh relablty. Thus, we must operate the dsks only n safe utlzaton zones. Fg. Queung model of RAID wth read workloads In the polcy we defned, the processor generates the read requests (See Fg that should be processed by the dsks. These read requests are queued up n a buffer and the utlzaton levels are calculated. These utlzaton levels are checked aganst the lmts of the safe utlzaton zone and then the dsk modes are changed accordngly. Ths approach s not only relable but also saves a sgnfcant amount of energy. Ths approach saves more energy for Travelstar dsks when compared wth Ultrastar dsks, because the energy levels of the dsk parameters lke spn up, spn down, actve power, dle power and etc. are much hgher for Ultrastar dsks when compared to Travelstar dsks. It should be noted here that the dsk parameters play a major role n the energy consumpton of the dsk. The rest of the paper s organsed as follows. In the next secton we dscuss the related work and motvaton. In secton 3, we descrbe the dsk falure model. Relablty aware power conservaton model s explaned n secton 4. In secton 5, we evaluate the performance RARE based on real world traces. Secton 6 concludes the paper wth the summary. 2. Related Work Extensve research has been carred out n developng energy effcent storage systems. Dynamc voltage scalng [3][9][3], dynamc power management[7], compler drected energy optmzatons [5][6] are some of the state of the art energy conservaton technques. Du el al. studed the dynamc voltage scalng technque wth a real-tme garbage collecton mechansm to reduce the energy dsspaton of flash memory storage systems [8]. A dynamc spn down technque for moble computng was proposed by Helmbold et al. [4]. A mathematcal model for each Dynamc Voltage Scalng - enabled system s bult and ther potental n energy reducton s analyzed by Ln Yuan and Gang Qu [3]. Carrera et al. [7] proposed four approaches to conservng dsk energy n hghperformance network servers and concluded that the fourth approach, whch uses multple dsk speeds, s the one that can actually provde energy savngs. An energy savng polcy named eraid [2] for conventonal dsk based RAID- systems usng redundancy s gven by L et al. Energy effcent dsk layouts for RAID- systems [2] have been proposed by Lu et al. Yue et al. nvestgated the memory energy effcency of hgh-end data servers used for supercomputers [0]. Son et al. proposed and evaluated a compler-drven approach to reduce dsk power consumpton of array-based scentfc applcatons executng on parallel archtectures [5][6][9]. Pnhero et al presented falure statstcs and analyzed the correlaton between falures and several parameters generally beleved to mpact longevty [5]. Four causes of varablty and an explanaton on how each s responsble for a possble gap between expected and measured drve relablty, are elaborated by Elerath and Shah [8]. Dempsey, a dsk smulaton envronment that ncludes accurate modelng of dsk power consumpton s presented by Zedlewsk et al []. They also demonstrated that dsk power consumpton can be smulated both effcently and accurately. Optmal power management polces for a laptop hard dsk are obtaned wth a system model that can handle non-exponental nter-arrval tmes n the dle and the sleep states [7]. Schroeder and Gbson presented and analyzed the feld-gathered dsk replacement data from fve systems n producton use at three organzatons []. They found evdence that falure rate s not constant wth age, and that there was a sgnfcant nfant mortalty effect. The nfant falures had a sgnfcant early onset of wear-out degradaton. Sgnfcant levels of correlaton between falures, ncludng autocorrelaton and long-range dependence, were also found. Gurumurth et al. [4] provded a new approach called DRPM to modulate dsk speed (RPM dynamcally, whch gves a practcal mplementaton to explot ths mechansm. They showed that DRPM can provde sgnfcant energy savngs wthout heavly 2
compromsng performance. Rost et al. presented a formal model of the behavor of CPU and I/O nteractons n scentfc applcatons, from whch they derved varous formulas that characterze applcaton performance [6]. All of the prevously mentoned work ether concentrated on the power conservaton or on the dsk relablty. Not many researchers address both energy effcency and relablty. It s very mportant for a data dsk to be very relable, whle consumng less power. The mportance of energy effcency and relablty, and the lack of research of ther relatonshp, motvates the research conducted n ths paper. 3. Dsk Falure Model Let Z+ be the set of postve ntegers. Wthout loss of generalty, we consder a workload condton where there are m Z+ dsk I/O phases. The utlzaton U of the th ( m I/O phase s a constant that can be straghtforwardly derved from the dsk I/O requrements of data-ntensve tasks runnng wthn the th I/O phase. Let φ be the number of data-ntensve tasks runnng n the th phase. Let λ j, j φ, denote the arrval rate of dsk request submtted by the jth data-ntensve task to the dsk system. Let s j, j φ be the average data sze of dsk requests of the jth task. The dsk I/O requrement of the jth task n the th phase s a product of the task s request arrval rate and the average data sze of dsk requests ssued by the task,.e., λ j sj. The accumulatve dsk I/O requrements R, measured n terms of MByte/Sec., of all the tasks runnng n the th phase can be wrtten as: φ ( λj j R s j. ( Note that R n Eq. ( can be envroned as accumulatve data amount access per tme unt. The utlzaton of a dsk system wthn a gven I/O phase equals to the rato the accumulatve dsk requrement R and the bandwdth of the dsk system. Thus, the utlzaton U of the th I/O phase can be expressed as U R B dsk φ ( λj sj j B dsk, (2 where B dsk s the bandwdth of the dsk system. The utlzaton U of the dsk system durng the m I/O phases s the weghted sum of the dsk utlzaton of all the m phases. Thus, the utlzaton U s expressed by Eq. (3 as follows: U m R R U B dsk φ ( λj sj j m φ ( λj sj j 2. (3 Gven I/O requrements of data-ntensve tasks ssung dsk request to a dsk system, one can leverage the above model to quantfy utlzaton of the dsk system. To determne the falure rate for a gven utlzaton rate, we took the ponts from the google study [5] and used the cubc splne nterpolaton method to approxmate annual falure rate of a dsk wth certan utlzaton and age. The dsk falure model can be modeled as an n+dmensonal r vector θ [ θ 0, θ 2,..., θ n ], where θ, 0 n, s the vector θ ( x, y that captures the correlatons between utlzaton x and dsk falure rate y. To develop the dsk falure rate model, we have to determne n+dmensonal vector θ r. Thus, gven the value of utlzaton x, one can make use of the falure rate model to calculate the falure rate y n component θ of vector θ r. To acheve ths goal, we adopted the cubc splne nterpolaton to generate falure rates n+dmensonal vectorθ r such that t results n a smooth curve for a gven dsk age. Cubc splne generates y f(x ax3 +bx2+cx+d That s a cubc functon for each nterval s used to plot the annual falure rate percentles for dsk systems of ages 2 years (see Fg. 2. and as functons of dsk utlzatons. Fg. 2. AFR for 2 year old dsk wth respect to utlzaton. Interestngly, results shown n Fgs. 2- contradct the fndngs of the prevous studes that ndcate that lower utlzaton levels produce lower falure rates whereas 3
hgher utlzaton levels correspond to hgher falure rates. The trend of a dsk annual falure rate as the dsk utlzaton grows s dfferent from that reported n the lterature. Specfcally, our dsk falure rate model bult from real-world data of dsk falures suggests that the probablty that a dsk fals under ether very low or very hgh utlzaton levels s very hgh. In contrast, when the utlzaton levels stay wthn a certan range (e.g., between 20% and 60% when the dsk age s 2 years old, see Fg. 2, the probablty of falures s far smaller than a specfed threshold (e.g., smaller than 2%. We term ths range of utlzaton levels as the safe utlzaton zone, whch largely depends on dsk ages. Gven a dsk system, ts safe utlzaton zone s a functon of the dsk s age and the specfed low falure rate threshold. 4. Relablty-aware power conservaton model RAID s popular and s wdely used for dsk drves. RAID s mplemented wth a mnmum of two dsks, whch are the prmary and back dsks. Intally the data s stored to the prmary dsk and then t s mrrored to the backup dsk. Ths mrrorng helps to recover the data when there s a falure n the prmary dsk. It also helps to ncrease the performance of the RAID system by sharng the workload between the dsks. We consdered RAID for all of our experments. The processor n the system generates the I/O stream, whch s queued to the buffer. The utlzaton of the dsk s calculated usng the request arrval rate. Please refer to Secton 3 for detals of the descrpton for the dsk utlzaton model. It should be noted that all requests here are consdered as read requests. At any gven pont of tme the dsks can be n the followng three states. State : Both the dsks n sleep mode State 2: Prmary dsk actve and backup dsk n sleep mode State 3: Both the dsks n actve mode and share the load. Let us consder that the dsks are n state at the begnnng. Once the utlzaton s calculated, t s compared wth the safe utlzaton zone range. If the calculated value falls below the range then dsks stay n state. If the calculated value s wthn the range, then the prmary dsk s made actve whle the backup dsk contnues to stay n the sleep mode. Ths represents a transton to state 2. If the calculated value s beyond the range then both the dsks are made actve and both of them share the load, whch corresponds to state 3. Transton of states from one power mode to another nvolves dsk spn up and/or spn down. The dsk spn ups and spn downs also consume a lot of energy. 5. Performance Evaluaton RAID s used n our experments. RAID uses a mnmum of two dsks, one as a prmary and one as the backup. We conducted the experments on three types of dsks from IBM. The expermental results are compared aganst two tradtonal state of the art methods. In the frst method, load balancng, both the dsks are always made actve. Load balancng acheves very hgh performance because both the dsks share the load. The second method, tradtonal method, s where the prmary dsk s made always actve and the backup dsk s kept n sleep mode. The backup dsk s made actve only when the utlzaton of the prmary dsk exceeds 00%, also known as saturaton. In what follows, we term our approach as RAREE (Relablty aware energy effcent approach. Fg. 3. Energy consumed by a 2 year old dsk The expermental data generated from the smulatons s plotted n fgure 3. Fgure 3 represents the energy consumed by the 2 year old dsk respectvely. In the above fgure RAREE represents the algorthm that s presented n ths paper. RAREE s compared aganst load balancng and the tradtonal method. From fg 3 t s observed that for the IBM 36Z5 dsk the power consumed by RAREE falls n between the load balancng and tradtonal technques. Even for the IBM 73LZX the trend s smlar, but the dfference n values s not as hgh as IBM 36Z5. For the IBM 40GNX the power consumed by RAREE s smaller than the tradtonal and load balancng power consumpton values because dsk spn up and spn down values are much smaller for the IBM 40GNX when compared wth the other two dsks. It should be observed that the dsk spn down and dsk spn up values play a vtal role n the energy consumpton. 4
Fg. 4. Energy consumed by a IBM 40GNX dsk wth dfferent ages Fg 4 shows the performance of RAREE on Travelstar dsks of dfferent ages. It can be observed from fgure that RAREE consumes less energy when compared to tradtonal and load balancng. defntely less than the two exstng technques, because RAREE makes the dsks go to sleep mode as soon there are no requests unlke the other technques. When dle power s changed unlke the actve power the energy consumed by the RAREE agan falls between the two technques. Ths s because RAREE makes the system go to sleep mode very often dependng on the condtons. It should be observed here though the RAREE consumed a bt hgher energy than tradtonal t can be neglected as we are achevng relablty. Fg. 5. Impact of spn up power on energy Fg 5 shows the effects on energy when the spn up energy s vared for a 2 year old dsk. The RAREE energy consumpton falls n between tradtonal and load balancng technques. Though the energy consumed by RAREE s a lttle hgher than tradtonal technque, here we are also ganng good amount relablty. Fg. 7. Relablty vs. dsk ages Fg 7 s a very mportant graph here t shows the relablty n terms of annual falure rate percentle. It can be observed from the graph that RAREE acheves a very hgh relablty when compared to load balancng and tradtonal. Only one bar s shown for load balancng and tradtonal technques because both have the same relablty levels as they don t pay specal attenton to relablty. We also found an nterestng observaton that when RAREE s appled to IBM40GNX, whch s a travelstar, t defntely consumes much less energy than the other two Ultrastars, whch are hgh performance dsks. Ths makes t clear that RAREE gves best results when t s used on moble dsks nstead of hgh performance dsks. Ths doesn t lmt the usage of RAREE to moble dsks because though Ultrastar consumes a lttle more energy than tradtonal technque we stll get a good relablty at a margnal cost of energy. Smulaton results prove that on an average roughly 20% of energy can be saved when RAREe s used nstead of load balancng. When RAREe s used nstead of the tradtonal method an excess of 3% of energy s saved, t s not a very sgnfcant amount but along wth a very lttle energy savng we are also achevng hgh relablty whch makes t sgnfcant. 6. Summary Fg. 6. Impact of actve power on energy Fg 6 shows the change n energy as the actve power s vared. Here RAREE energy consumpton s Although an array of energy conservaton schemes have been proposed for dsk systems, most energy conservaton technques are mplemented at cost of 5
dsk relablty. In order to solve ths problem, we frst buld a model to quantfy the relatonshp among the dsk age, utlzaton, and falure probabltes. In ths study, we focused on mrrorng dsk systems, where data sets are mrrored to backup dsks from prmary dsks. Tradtonal methods wake up the backup dsks when the utlzaton of the prmary dsks exceeds 00 percent. Load balancng technque keeps both the dsks always actve to balance the load between prmary and backup dsks to acheve hgh performance. These two methods consume a massve amount of energy. Hence, we amed at developng a mechansm to reduce the energy consumpton where we determne the safe utlzaton levels for the dsks to operate wth mnmum probablty falure rates whle conservng energy. After proposng a novel concept of safe utlzaton zone where energy of the dsk can be conserved wthout degradng relablty, we desgned and mplemented a utlzaton control mechansm to mprove both relablty and energy effcency of dsk systems. The utlzaton control mechansm ensures that dsk drves are operated n safe utlzaton zones to mnmze the probablty of dsk falure. We ntegrate the energy consumpton technque that operates the dsks at dfferent power modes wth our proposed relablty approach. Expermental results show that our approach can sgnfcantly mprove relable whle achevng hgh energy effcency for dsk systems. Acknowledgements The work reported n ths paper was supported by the US Natonal Scence Foundaton under Grants No. CCF-074287, No. CNS-0757778, No. CNS-083502, No. OCI-0753305, and No. DUE-062307, and Auburn Unversty under a startup grant. References [] B. Schroeder, and G.A. Gbson, Dsk falures n the real world: what does an MTTF of,000,000 hours mean to you? Proc.5th Conf. on USENIX Conf. on Fle and Storage Technologes, vol.5, San Jose, CA, Feb 2007. [2] D. L, and J. Wang, "Conservng Energy n RAID Systems wth Conventonal Dsks," Proc. 3rd Int'l Workshop on Storage Network Archtecture and Parallel I/Os, 2005. [3] D. Zhu, R. Melhem, and D. Mosse, "The effects of energy management on relablty n real-tme embedded systems", Proc. IEEE/ACM Int'l conf. Computer-aded desgn, pp. 35-40, 2004. [4] D.P. Helmbold, D.D. Long, and B. Sherrod, "A dynamc dsk spn-down technque for moble computng," Proc. 2nd annual Int'l Conf. on Moble computng and networkng, ACM New York, USA, 996, pp. 30-42. [5] E. Pnhero, W. Weber, and L. A. Barroso, Falure trends n a large dsk drve populaton, Proc. 5th Conf. on USENIX Conf. on Fle and Storage Technologes, vol.5, San Jose, CA, Feb 2007. [6] E. Rost, G. Serazz, E. Smrn and M.S. Squllante, "Models of Parallel Applcatons wth Large Computaton and I/O Requrements," IEEE Trans. Software Engneerng,vol. 28, no.3, pp. 286-307. [7] E. V. Carrera, E. Pnhero, and R. Banchn, "Conservng Dsk Energy n Network Servers," Proc. 7th Int'l Conf. on Supercomputng, ACM Press, pp. 86-97. [8] J. G. Elerath and S. Shah, Server class dsk drves: how relable are they, IEEE Relablty and Mantanablty Symp., pp. 5-56, Jan 2004. [9] J. Mao, C. G. Cassandras, Q. Zhao, Optmal Dynamc Voltage Scalng n Energy-Lmted Nonpreemptve Systems wth Real-Tme Constrants, IEEE Trans. Moble Computng, vol.6,jun 2007,pp. 678-688. [0] J. Yue, Y. Zhu, and Z. Ca, "Evaluatng Memory Energy Effcency n Parallel I/O Workloads", Proc. IEEE Internatonal Conf. on Cluster Computng, Austn, Texas, 2007, pp. 2-30 [] J. Zedlewsk, S. Sobt, N. Garg, F. Zheng, A. Krshnamurthy, and R. Wang, Modelng Hard-Dsk Power Consumpton, Proc. 2nd USENIX Conf on Fle and Storage Technologes, San Francsco, CA, March 2003. [2] L. Lu, P. Varman, and J. Wang, "DskGroup: Energy Effcent Dsk Layout for RAID Systems," Int'l Conf. on Networkng, Archtecture, and Storage, IEEE Press, pp. 233-242. [3] L. Yuan, and G. Qu, "Analyss of energy reducton on dynamc voltage scalng-enabled systems," IEEE Tran. Computer-Aded Desgn of Integrated Crcuts and Systems, vol. 24, no. 2, pp. 827-837. [4] S. Gurumurth, A. Svasubramanam, M. Kandemr, and H. Franke, DRPM: dynamc speed control for power management n server class dsks, Proc. 30th Int l Symp. Computer Archtecture, pp. 69-8, May 2003. [5] S. Son, G. Chen, M. Kandemr, and A. Choudhary," Exposng dsk layout to compler for reducng energy consumpton of parallel dsk based systems," Proc. ACM SI Symp. Prncples and practce of parallel programmng, pp. 74-85. [6] S. W. Son, G. Chen, O. Ozturk, M. Kandemr, and A. Choudhary, "Compler-Drected Energy Optmzaton for Parallel Dsk Based Systems," IEEE Trans. Parallel and Dstrbuted Systems, vol. 8, no.9, 2007, pp. 24-257. [7] T. Smunc, L. Benn, and G. De Mchel, "Dynamc Power Management of Laptop Hard Dsk," Proc. Desgn Automaton and Test, Europe, 2000. [8] Y. Du, J. Dong, and M. Ca, "Dynamc Voltage Scalng of Flash Memory Storage Systems for Low-Power Real- Tme Embedded Systems," Proc. Second Int'l Conf.Embedded software and systems, 2005, pp. 52-57. 6