The growing cost of tuning and managing computer systems is leading to out-sourcing commercial services hosting centers. These centers provision thousands dense servers within a relatively small real-estate in order host the applications/services different customers who may have been assured by service-level agreement (SLA). Power consumption these becoming serious concern design operation effects high power manifest not only costs spent designing effective cooling ward off generated heat,...

Since the electricity bill of a data center constitutes significant portion its overall operational costs, reducing this has become important. We investigate cost reduction opportunities that arise by use uninterrupted power supply (UPS) units as energy storage devices. This represents deviation from usual these devices mere transitional fail-over mechanisms between utility and captive sources such diesel generators. consider problem opportunistically using to reduce time average electric in...

In this paper, we explore the possibility of using STT-RAM technology to completely replace DRAM in main memory. Our goal is make performance comparable while providing substantial power savings. Towards goal, first analyze and energy STT-RAM, then identify key optimizations that can be employed improve its characteristics. Specifically, partial write row buffer bypass, show memory significantly improved. experiments indicate an optimized, equal capacity provide memory, with average 60%...

A large portion of the power budget in server environments goes into I/O subsystem - disk array particular. Traditional approaches to management involve completely stopping rotation, which can take a considerable amount time, making them less useful cases where idle times between requests may not be long enough outweigh overheads. This paper presents new approach called DRPM modulate speed (RPM) dynamically, and gives practical implementation exploit this mechanism. Extensive simulations...

As the complexity of distributed computing systems increases, management tasks require significantly higher levels automation; examples include diagnosis and prediction based on real-time streams computer events, setting alarms, performing continuous monitoring. The core autonomic computing, a recently proposed initiative towards next-generation IT-systems capable 'self-healing', is ability to analyze data in predict potential problems. goal avoid catastrophic failures through prompt...

The growing computational and storage needs of several scientific applications mandate the deployment extreme-scale parallel machines, such as IBM’s BlueGene/L which can accommodate many 128K processors. One challenges when designing deploying these systems in a production setting is need to take failure occurrences, whether it be hardware or software, into account. Ear- lier work has shown that conventional runtime fault- tolerant techniques periodic checkpointing are not effective emerging...

The growing complexity of hardware and software mandates the recognition fault occurrence in system deployment management. While there are several techniques to prevent and/or handle faults, continues be a need for an in-depth understanding errors failures their empirical statistical properties. This can help evaluate effectiveness different improving availability, addition developing new solutions. In this paper, we analyze properties from network nearly 400 heterogeneous servers running...

Workload consolidation is very attractive for cloud platforms due to several reasons including reduced infrastructure costs, lower energy consumption, and ease of management. Advances in virtualization hardware software continue improve resource isolation among consolidated workloads but a particular form interference yet see commercially widely adopted solution - the shared processor caches. Existing solutions handling cache require new features, extensive changes, or reduce achieved...

Datacenter power consumption has a significant impact on both its recurring electricity bill (Op-ex) and one-time construction costs (Cap-ex). Existing work optimizing these relied primarily throttling devices or workload shaping, with performance degrading implications. In this paper, we present novel knob of energy buffer (eBuff) available in the form UPS batteries datacenters for cost optimization. Intuitively, eBuff stores during "valleys" - periods lower demand, which can be drained...

Recent advances in software and architectural support for server virtualization have created interest using this technology the design of consolidated hosting platforms. Since enables easier faster application migration as well secure co-location antagonistic applications, higher degrees consolidation are likely to result such virtualization-based platforms (VHPs). We identify a key shortcoming existing virtual machine monitors (VMMs) that proves be an obstacle operating platforms, Internet...

Energy storage - in the form of UPS units a datacenter has been primarily used to fail-over diesel generators upon power outages. There recent interest using these Storage Devices (ESDs) for demand-response (DR) either shift peak demand away from high tariff periods, or shave allowing aggressive under-provisioning infrastructure. All such prior work only considered single/specific type ESD (typically re-chargeable lead-acid batteries), and employed them at single level delivery network....

Power dissipation has become one of the most critical factors for continued development both high-end and low-end computer systems. We present a complete system power simulator, called SoftWatt, that models CPU, memory hierarchy, low-power disk subsystem quantifies behavior application operating system. This tool, built on top SimOS infrastructure, uses validated analytical energy to identify hotspots in components, capture relative contributions user kernel code profile, power-hungry...

The growing cost of tuning and managing computer systems is leading to out-sourcing commercial services hosting centers. These centers provision thousands dense servers within a relatively small real-estate in order host the applications/services different customers who may have been assured by service-level agreement (SLA). Power consumption these becoming serious concern design operation effects high power manifest not only costs spent designing effective cooling ward off generated heat,...

The last line of defense in the cache hierarchy before going to off-chip memory is very critical chip multiprocessors (CMPs) from both performance and power perspectives. We investigate different organizations for this (assumed be L2 article) towards reducing accesses. evaluate trade-offs between private address-interleaved shared designs, noting their individual benefits drawbacks. possible imbalance demands across CPUs favors a organization, while interference these can favor organization....

Datacenters are facing increasing pressure to cap their carbon footprints at low cost. Recent work has shown the significant environmental benefits of using renewable energy for datacenters by supply-following techniques (workload scheduling, geographical load balancing, etc.) However, all such prior only considered on-site generation when numerous other options also exist, which may be superior renewables many datacenters. Alternative ways incorporate into overall portfolio include:...

Current capacity planning practices based on heavy over-provisioning of power infrastructure hurt (i) the operational costs data centers as well (ii) computational work they can support. We explore a combination statistical multiplexing techniques to improve utilization hierarchy within center. At highest level hierarchy, we employ controlled underprovisioning and over-booking needs hosted workloads. lower levels, introduce novel notion soft fuses flexibly distribute provisioned among...

Datacenters spend $10-25 per watt in provisioning their power infrastructure, regardless of the watts actually consumed. Since peak needs arise rarely, infrastructure for them can be expensive. One can, thus, aggressively under-provision assuming that simultaneous draw across all equipment will happen rarely. The resulting non-zero probability emergency events where exceed provisioned capacity, however small, mandates graceful reaction mechanisms to cap instead leaving it disruptive circuit...

NAND flash-based solid-state drives (SSDs) are increasingly being deployed in storage systems at different levels such as buffer-caches and even secondary storage. However, the poor reliability performance offered by these SSDs for write-intensive workloads continues to be their key shortcoming. Several solutions based on traditionally popular notions of temporal spatial locality help reduce write traffic SSDs. another form - value has remained completely unexplored. Value implies that...

Unlike the use of DRAM for caching or buffering, certain idiosyncrasies SSDs make their integration into existing systems non-trivial. Flash memory suffers from limits on its reliability, is an order magnitude more expensive than HDD, and can sometimes be as slow HDD (due to excessive garbage collection (GC) induced by high intensity random writes). Given these trade-offs between HDDs in terms cost, performance, lifetime, current consensus among several storage experts view not a replacement...

While there have been several studies and proposals for energy conservation CPUs peripherals, optimization techniques selective operating mode control of DRAMs not fully explored. It has shown that as much 90% overall system (excluding I/O) is consumed by the DRAM modules, serving a good candidate optimizations. Further; technology also matured to provide low modes (power modes), making it an opportunistic moment conduct exploring potential benefits techniques. This paper conducts in-depth...

The importance of the Translation Lookaside Buffer (TLB) on system performance is well known. There have been numerous prior efforts addressing TLB design issues for cutting down access times and lowering miss rates. However, it was only recently that first exploration [26] prefetching entries ahead their need undertaken a mechanism called Recency Prefetching proposed. large body literature caches, not clear how they can be adapted (or if are different) TLBs, suited prefetching, compare with...

Previous work on DRAM power-mode management focused hardware-based techniques and compiler-directed schemes to explicitly transition unused memory modules low-power operating modes. While require extra logic keep track of references make decisions about future mode transitions, can only a single application at time demand sophisticated program analysis support. In this work, we present an system (OS) based solution where the OS scheduler directs power transitions by keeping module accesses...

For a peer-to-peer (P2P) system holding massive amount of data, efficient semantic based search for resources (such as data or services) is key determinant to its scalability. This work presents the design an overlay network, namely small world (SSW), that facilitates in P2P systems. SSW on three innovative ideas: 1) network; 2) clustering; 3) dimension reduction. Peers are clustered according semantics their local and self-organized network. To address maintenance issue high dimensional...

The growing computational and storage needs of several scientific applications mandate the deployment extreme-scale parallel machines, such as IBM's BlueGene/L, which can accommodate many 128K processors. In this paper, we present our experiences in collecting filtering error event logs from a 8192 processor BlueGene/L prototype at IBM Rochester, is currently ranked #8 Top-500 list. We analyze collected machine over period 84 days starting August 26, 2004. perform three-step algorithm on...