Presentation on theme: "13th September, 2005CSE 598B: Fall2005 Presentation 1 Automated administration for storage system Presentation by Amitayu Das."— Presentation transcript:
13th September, 2005CSE 598B: Fall2005 Presentation 1 Automated administration for storage system Presentation by Amitayu Das
13th September, 2005CSE 598B: Fall2005 Presentation 2 Introduction Major challenges in storage management – System design and configuration (device management) – Capacity Planning (space management) – Performance tuning (performance management) – High Availability (availability management) – Automation (all of the above, in a self- managing manner)
13th September, 2005CSE 598B: Fall2005 Presentation 3 Motivation Large disk arrays and networked storage lead to huge storage capacities and high bandwidth access to facilitate consolidated storage systems. Enterprise-scale storage systems contain hundreds of host computers and storage devices and up to tens of thousands of disks. Designing, deploying and runtime management of such systems lead to huge cost (often higher than procuring cost)… Look at the problems in greater details …
13th September, 2005CSE 598B: Fall2005 Presentation 7 Designing problem Given a pool of resources and workload, determine appropriate choice of devices, configure them and assign the workload to the configured storage. Solution is not straight-forward because, – Huge size of system and thousands of design choices and many choices have unforeseen circumstances. – Personnel with detailed knowledge of applications’ storage behavior are in short supply and hence, are quite expensive. – Design process is tedious and complicated to do by hand, usually leading to solutions that are grossly over-provisioned, substantially under-performing or, in the worst case, both. – Once a design is in place, implementing it is time-consuming, tedious and error-prone. – A mistake in any of these steps is difficult to identify and can result in a failure to meet the performance requirements.
13th September, 2005CSE 598B: Fall2005 Presentation 9 Application System design and assignment problem Application Assignment engine Storage System Storage System configuration Storage device abilities Workload Workload requirements
13th September, 2005CSE 598B: Fall2005 Presentation 10 Initial system design Problem: convert workloads, business needs and device characteristics into assignment of stores and streams to devices One approach: constraint-based multi-dimensional bin-packing Sample constraints: # of device = 1 – - Sum of store sizes capacity – - Sum of stream utilizations 1.0 Sample objective functions: – - Minimize cost – - Balance load Req. size Capacity I/O rate How many drives? Holding which data?
13th September, 2005CSE 598B: Fall2005 Presentation 11 Initial system design –> disk arrays Problem: – extending the single disk solution to disk arrays – The space of array designs is potentially huge: LUN sizes and RAID levels, stripe unit sizes, disks in LUNs More work needed before the solver can run
13th September, 2005CSE 598B: Fall2005 Presentation 12 Minerva Control flow. The array designer is called as a subroutine by allocator. Minerva’s role in storage system life cycle. Input and output are shown.
13th September, 2005CSE 598B: Fall2005 Presentation 14 Merits/demerits Merits: – Reasonable automation Demerits: – Requires accurate models of workloads, performance requirements, and devices – Address only the mechanisms, not the policy
13th September, 2005CSE 598B: Fall2005 Presentation 16 System redesign/reconfiguration Running System Reconfigured System new application added new users added system load increases hardware/software upgraded device fails new storage arrives performance tuning Events triggering redesign/reconfiguration
13th September, 2005CSE 598B: Fall2005 Presentation 18 Hippodrome Two objectives: – The automated loop must converge on a viable design that meets the workload’s requirements without over- or under- provisioning. – It must converge to a stable final system as quickly as possible, with as little as input from its users.
13th September, 2005CSE 598B: Fall2005 Presentation 20 Issues in system design and allocation What optimization algorithms are most effective? What optimization objectives and constraints produce reasonable designs? – ex: cost of reconfiguring system What's the right part of the storage design space to explore? – ex: RAID level vs. stripe unit size vs. cache management parameters What are reasonable general guidelines for tagging a store's RAID level? What (other) decompositions of the design and allocation problem are reasonable? How to generalize system design? – for SAN environment – for host and applications
13th September, 2005CSE 598B: Fall2005 Presentation 21 Issues in reconfiguration How to do system discovery? – e.g., existing state, presence of new devices – Dealing with inconsistent information – In a scalable fashion How to abstractly describe storage devices? – For system discovery output – For input to tools that perform changes How to automate the physical redesign process? – e.g., physical space allocation etc. Events trigger redesign decision – – How do we decide when to reconfigure? Reconfiguration inputs: – current system configuration/assignment – desired system configuration/assignment
13th September, 2005CSE 598B: Fall2005 Presentation 23 Administration and organization Administrative interface Supervisors Administrative assistants Data access and storage Routers Workers
13th September, 2005CSE 598B: Fall2005 Presentation 24 Merits Simpler storage administration – Data protection – Performance tuning – Planning and deployment – Monitoring and record-keeping – Diagnosis and repair
13th September, 2005CSE 598B: Fall2005 Presentation 25 Demerits The proposed solution is too simplistic to handle the issues raised. Authors have provided solution from a high-level viewpoint, but the solution is not complete in any sense. The implementation and evaluation is not convincing enough. All the aspects of “self-*” has not been addressed as claimed.
13th September, 2005CSE 598B: Fall2005 Presentation 27 Runtime management problem Often, enterprise customers outsource their storage needs to data centers. At data centers, different workload /application /services share the underlying storage infrastructure. Sharing (of disk drives, storage caches, network links, controllers etc.) can lead to interference between the users/applications leading to possible violations in performance-based QoS guarantees. To prevent that, data centers needs to insulate the users from each other – virtualization.
13th September, 2005CSE 598B: Fall2005 Presentation 28 Need for virtualization At data centers, many different enterprise servers that support different business processes, such as, Web servers, file servers, database serves may have very different performance requirements on their backend storage server. Sophisticated resource allocation and scheduling technology is required to effectively isolate these logical storage servers as if they are separate physical storage servers. Storage Virtualization refers to the technology that allows creation of a set of logical storage devices from a single physical storage structure.
13th September, 2005CSE 598B: Fall2005 Presentation 30 Dimensions of virtualization Commercial storage virtualization systems are rather limited because they can virtualize storage capacity. However, from the standpoint of storage clients or enterprise servers, the virtual storage devices are desired to be as tangible as physical disks. Need to virtualize efficiently any standard attribute associated with a physical disk, such as capacity, bandwidth, latency, availability etc.
13th September, 2005CSE 598B: Fall2005 Presentation 31 Hardware Organization Storage manager Storage server Disk array Kernel Client Application Storage Clerk Kernel Client Application Storage Clerk Storage server Disk array Storage server Disk array Control mesg Data/cmds Gigabit network Object interface client File interface
13th September, 2005CSE 598B: Fall2005 Presentation 33 References Hippodrome: running circles around storage administration. Eric Anderson et. al., FAST ’02, pp. 175-188, January 2002. Minerva: an automated resource provisioning tool for large-scale storage systems. G. Alveraz et. al., ACM Transactions on Computer Systems 19 (4): 483-518, November 2001 Ergastulum: quickly finding near-optimal storage system designs. Eric Anderson et. al., Technical Report from HP Laboratories. Disk Array Models in Minerva. Arif Merchant et. al., Technical Report, HP Laboratories. Self-* Storage: Brick-based Storage with Automated Administration. G. Ganger et. al., Technical report,2003
13th September, 2005CSE 598B: Fall2005 Presentation 34 References SIGMETRICS ’00 Tutorial, HP Laboratories. Optimization algorithms – Bin-packing Heuristics [Coffman84] – Toyoda Gradient [Toyoda75] – Simulated Annealing [Drexl88] – Relaxation Approaches [Pattipati90, Trick92] – Genetic Algorithms [Chu97] Multidimensional Storage Virtualization. Lan Huang et. al., SIGMETRICS ’04, New York, June 2004. An Interposed 2-Level I/O Scheduling Framework for Performance Virtualization. J. Zhang et. al., SIGMETRICS ’05 Efficiency-aware disk scheduler: – - Cello, Prism, YFQ