Distributed Organisation Seminar Talk: Information Grouping Framework For Energy-Efficiency Inwards Distributed Storage Systems
My inquiry grouping in addition to Tevfik's inquiry group come across jointly for a weekly distributed systems. This gives our students a hazard to laissez passer talks near electrical flow projection in addition to larn feedback for improvement inwards a friendly setting.
In this week's seminar, Luigi presented his inquiry on edifice energy-efficient file systems. I was initially skeptical near energy-efficiency every bit a inquiry topic. Academicians similar to piece of work on things that they tin quantify in addition to improve, in addition to then I was thinking that energy-efficiency inwards distributed storage was an opportunistic inquiry problem, rather than a real-world problem. Turns out, I couldn't live on whatever to a greater extent than wrong: information technology companies pass $10 billions every yr on unloosen energy consumption (This is 3% of entire expenditure of US!). $3.5 billion of that $10 billion is unloosen energy expenditure is due to the storage systems.
Dynamic ability management (DPM) is the principal machinery for unloosen energy saving at the storage systems. DPM basically way plow the disk off if you're non using it. An idling disk spends unloosen energy because it is yet rotating, in addition to this mechanic displace which burns energy. But turning a disc off is non easy. It takes 10s of seconds to halt in addition to outset difficult disk, in addition to the unloosen energy usage spikes at these transition points. This makes the work into an optimization problem. When is it beneficial to plow the disk off? How tin you lot create gaps long plenty to plow off the disk?
The literature discusses the next DPM-enabling techniques for energy-saving inwards storage systems. Most of these techniques prescribe information access locality improvements.
1) Memory in addition to disk caching: Caching is non exclusively practiced for providing low-latency merely also inwards roughly cases practiced for saving energy. If nosotros tin piece of work cache to response instead of turning on the disk, nosotros tin laissez passer the disk to a greater extent than fourth dimension to sleep. But what should live on the cache size? If it is besides small, information won't fit, this won't supply much/any saving. If it is besides large, the cache itself may eat to a greater extent than unloosen energy than it saves.
2) Diverting accesses: Data is stored redundantly, in addition to then this gives us the chance to spin downwards roughly redundant disks past times diverting the accesses to the already active/hot ones. Unsurprisingly, in that location is a tradeoff of increased latency inwards doing so. By limiting concurrency/parallelism you lot increment latency of replies. (Is energy-efficiency versus latency a key tradeoff inwards distributed storage?) Maybe, by offering well-drafted SLA agreements to the clients, it is possible to laissez passer incentive to the customer for trading unloosen energy efficiency for slightly increased latency.
3) Popular information clustering: This technique prescribes organizing the disk storage based on the previously observed access locality of data. So if a disk is hot, it is probable to remain hot, in addition to if a disk gets cold, it is probable to remain mutual coldness in addition to it tin sleep.
I gauge in that location also could live on orthogonal techniques if you lot don't ask to serve requests inwards real-time. For those cases you lot receive got the chance to batch-schedule accesses.
Luigi is working on a hybrid of these techniques to supply every bit much energy-efficiency every bit possible. I wouldn't receive got idea energy-efficiency for distributed storage could live on this interesting. There mightiness fifty-fifty live on a duet distributed algorithms work hither that I would enjoy.
In this week's seminar, Luigi presented his inquiry on edifice energy-efficient file systems. I was initially skeptical near energy-efficiency every bit a inquiry topic. Academicians similar to piece of work on things that they tin quantify in addition to improve, in addition to then I was thinking that energy-efficiency inwards distributed storage was an opportunistic inquiry problem, rather than a real-world problem. Turns out, I couldn't live on whatever to a greater extent than wrong: information technology companies pass $10 billions every yr on unloosen energy consumption (This is 3% of entire expenditure of US!). $3.5 billion of that $10 billion is unloosen energy expenditure is due to the storage systems.
Dynamic ability management (DPM) is the principal machinery for unloosen energy saving at the storage systems. DPM basically way plow the disk off if you're non using it. An idling disk spends unloosen energy because it is yet rotating, in addition to this mechanic displace which burns energy. But turning a disc off is non easy. It takes 10s of seconds to halt in addition to outset difficult disk, in addition to the unloosen energy usage spikes at these transition points. This makes the work into an optimization problem. When is it beneficial to plow the disk off? How tin you lot create gaps long plenty to plow off the disk?
The literature discusses the next DPM-enabling techniques for energy-saving inwards storage systems. Most of these techniques prescribe information access locality improvements.
1) Memory in addition to disk caching: Caching is non exclusively practiced for providing low-latency merely also inwards roughly cases practiced for saving energy. If nosotros tin piece of work cache to response instead of turning on the disk, nosotros tin laissez passer the disk to a greater extent than fourth dimension to sleep. But what should live on the cache size? If it is besides small, information won't fit, this won't supply much/any saving. If it is besides large, the cache itself may eat to a greater extent than unloosen energy than it saves.
2) Diverting accesses: Data is stored redundantly, in addition to then this gives us the chance to spin downwards roughly redundant disks past times diverting the accesses to the already active/hot ones. Unsurprisingly, in that location is a tradeoff of increased latency inwards doing so. By limiting concurrency/parallelism you lot increment latency of replies. (Is energy-efficiency versus latency a key tradeoff inwards distributed storage?) Maybe, by offering well-drafted SLA agreements to the clients, it is possible to laissez passer incentive to the customer for trading unloosen energy efficiency for slightly increased latency.
3) Popular information clustering: This technique prescribes organizing the disk storage based on the previously observed access locality of data. So if a disk is hot, it is probable to remain hot, in addition to if a disk gets cold, it is probable to remain mutual coldness in addition to it tin sleep.
I gauge in that location also could live on orthogonal techniques if you lot don't ask to serve requests inwards real-time. For those cases you lot receive got the chance to batch-schedule accesses.
Luigi is working on a hybrid of these techniques to supply every bit much energy-efficiency every bit possible. I wouldn't receive got idea energy-efficiency for distributed storage could live on this interesting. There mightiness fifty-fifty live on a duet distributed algorithms work hither that I would enjoy.
0 Response to "Distributed Organisation Seminar Talk: Information Grouping Framework For Energy-Efficiency Inwards Distributed Storage Systems"
Post a Comment