Kester Li, Toward a low power file system, Master's thesis, University of California at Berkeley, May 1994. 22  Home/Search   Document Details and Download   Summary   Related Articles   Check

....to show differences between systems performing different tasks such as the clients and servers of a distributed file system. Others have used kernel instrumentation to gain specific information about a machine. Ruemmler and Wilkes [86] were interested particularly in the disk s active data set. Li [53] instrumented MS DOS machines to gain information about augmenting cache behavior in those systems. Mummert and Satyanarayanan [67, 99] detail a distributed system kernel tracing facility. In this facility, machines will forward their tracing results onto a central machine for storage. In this ....

Li, K. Towards A Low Power File System. Tech. Rep. UCB:CSD-94-814, Department of Computer Science, University of California, Berkeley, May 1994.

....don t further improve energy consumption by much. For more coarse grain spin down (e.g. spin down after five minutes idle) Marsh et al. 94] found that increasing the cache size to beyond 16 MB still significantly improves power consumption, but not as much as fine grain spin down of the disk. Li 94] and [Marsh et al. 94] explore using flash memory as a second level cache between the DRAM cache and the disk. Both found that a 4 to 8 MB flash cache further improves the energy consumption of the disk subsystem by up to 50 . Moreover, Li 94] shows that a cache that prefetches whole files on ....

....but not as much as fine grain spin down of the disk. Li 94] and [Marsh et al. 94] explore using flash memory as a second level cache between the DRAM cache and the disk. Both found that a 4 to 8 MB flash cache further improves the energy consumption of the disk subsystem by up to 50 . Moreover, Li 94] shows that a cache that prefetches whole files on the first read improves over this by another 50 . 6.3 Flash Memory as a Disk Alternative Flash memory has been proposed as a possible alternative to hard disks. It has much lower power consumption and faster read response time, and like disks ....

[Article contains additional citation context not shown here]

Li, K. Towards a Low Power File System. Master's thesis, University of California, Berkeley, May 1994.

....into disk level operations, by associating a unique disk location with each file. An interesting aspect of the mac trace is the locality of write accesses: 36 of writes go to just one 1 Kbyte block and 24 of writes go to another. We used a dos trace collected by Kester Li at U.C. Berkeley [13], on IBM desktop PCs running Windows 3.1, also at file level. It includes deletions. The trace was similarly preprocessed. 10 Chapter 1 We used a disk level trace collected by Ruemmler and Wilkes on an hp personal workstation running HP UX [21] This trace include metadata operations, which the ....

....taking 20ms plus a value that is exponentially distributed with a mean of 3s. Though only the mac trace comes from a mobile environment, the two desktop traces represent workloads similar to what would be used on mobile computers, and have been used in simulations of mobile computers in the past [13, 14, 15]. Table 3 lists additional statistics for the nonsynthetic traces. 4.2 Simulator Our simulator models a storage hierarchy containing a buffer cache and nonvolatile storage. The buffer cache is the first level searched on a read and is the target of all write operations. The cache is ....

Kester Li. Towards a low power file system. Technical Report UCB/CSD 94/814, University of California, Berkeley, CA, May 1994. Masters Thesis.

....Memory One method to reduce the number of times a hard disk has to be spun up is to have a large amount of system memory (i.e. DRAM) This makes intuitive sense place the current working set in memory and there will be few page faults to cause the disk to spin up. Unfortunately a study by Li [17] has shown that having as little as 8 MBytes of additional DRAM can use up as much power as a constantly spinning hard disk. To confirm this rather surprising result, we did some calculations based on manufacturers data [6] and found that 8 MBytes of 60ns Extended Data Output (EDO) DRAM uses 2.8 ....

....potential performance hit. If a user could set (either at power up, or dynamically) the amount of memory to be powered down, it would offer a very straightforward way to trade off performance with laptop battery endurance. Since about 8 Mbytes of DRAM can use as much power as a spinning hard disk [17], the additional page faults (and subsequent drive spin ups) would be offset by the savings from having reduced DRAM. Intel has released a new Pentium PCI chipset (the 82430MX PCI chipset) that has suspend and standby modes which not only put the CPU in low power mode, but also restrict power to ....

Kester Li, Towards A Low Power File System, Technical Report UCB/CSD 94/814, University of California - Berkeley, Masters Thesis, May 1994.

....computer. There are three major methods that we will look at: ffl Pure mechanical control: Controlling the spin down of the hard disk [10, 4, 6] ffl Removing the disk: Using flash memory as the only permanent storage device [9] ffl Hybrid system: Using flash memory and the hard disk together [12, 11] The rest of this paper is organized as follows. We will first show the trend for the proportion of power consumption by each of the major subsystems. We will then briefly describe the types of memory currently used in computers and the advantages and shortcomings of the ones relevant to this ....

....threshold value, with a 19 increase in delays. Another trace with the same disk resulted in a 15 power reduction but a 150 increase in delays. It would be safe to say that the current Threshold Demand algorithm is inadequate for the future. A similar study was done by Li et al. in [11]. They discuss in detail the sleep mode that is available in modern laptop disks a summary of disk modes is presented in Figure 3. The DOS traces used by Li in their simulation are by far the longest and most comprehensive of all the simulations for mobile computers so far they spanned a ....

[Article contains additional citation context not shown here]

Kester Li, Towards A Low Power File System, Technical Report UCB/CSD 94/814, University of California - Berkeley, Masters Thesis, May 1994.

....(20 ) and the CPU (12 ) 6] There is seemingly little which can be done to minimize screen power consumption, beyond employing a screen saver and waiting for hardware improvements. Disk power consumption may be minimized by spinning down the disk when it has been inactive for several seconds; [3, 6, 7] have researched this topic. kinshuk csua.berkeley.edu. y chance cory.eecs.berkeley.edu. z halw cs.berkeley.edu. Supported by NDSEG Fellowship daah04 93 g 0267. In the future we may well see ubiquitous computing devices with neither disks nor conventional displays; and, for such devices, ....

K. Li, "Towards a low power file system," CS Tech. Report 94-814, University of California, Berkeley, May 1994.

....number of times a hard disk has to be spun up is to have a large amount of system memory (i.e. DRAM) This makes intuitive sense place the current working set in memory and there will be few page faults to cause the disk to spin up. Unfortunately this is not feasible in practice. A study by Li [17] has shown that having as little as 8 MBytes of additional DRAM can use up as much power as a constantly spinning hard disk. To confirm this rather surprising result, we did some calculations based on manufacturers data [5] and found that 8 MBytes of 60ns Extended Data Output (EDO) DRAM uses 2.8 ....

....possibly with performance degradation. If a user could set (either at power up, or dynamically) the amount of memory to be powered down, it would offer a method to tradeoff performance with laptop battery endurance. Since about 8 Mbytes of DRAM can use as much power as a spinning hard disk [17], the additional page faults (and subsequent drive spin ups) would be offset by the savings from having reduced DRAM. Intel has released a new Pentium PCI chipset (the 82430MX PCI chipset) that has suspend and standby modes which not only put the CPU in low power mode, but also restrict power to ....

Kester Li, Towards A Low Power File System, Technical Report UCB/CSD 94/814, University of California - Berkeley, Masters Thesis, May 1994.

....reducing cleaning costs. This report is also submitted to Journal of Systems and Software and is accepted for publications. # However, flash memory requires additional system support for erasure management because of the hardware characteristics [Caceres et al. 1993; Diper and Levy, 1993; Douglis et al. 1994; Intel, 1994; Intel, 1997; Kawaguchi et al. 1995; Wu and Zwaenepoel, 1994] shown in Table 1. Flash memory is partitioned into segments 2 defined by hardware manufacturers (e.g. 64 Kbytes or 128 Kbytes for Intel Series 2 Flash Memory Cards [Intel, 1994; Intel, 1997] and 512 bytes for SanDisk ....

....We used only the traces from personal workstation (hplajw) which was used primarily for electronic mail and document editing. Because the usage behavior of personal computers is likely to be similar to mobile computers, hplajw traces were often used in simulations of mobile computers [Douglis et al. 1994; Li, 1994; Li et al. 1994, Marsh et al. 1994] However, the disks (334 Mbytes) used in hplajw are much larger than flash memory. Therefore, the traces were preprocessed to map flash memory spaces before simulation. The original traces exhibit such high locality of reference that 71.2 of writes ....

[Article contains additional citation context not shown here]

Li, K., Towards a low power file system, Technical Report UCB/CSD 94/814, Masters Thesis, University of California, Berkeley, CA, May 1994.

....and the CPU (12 ) 4] There is seemingly little which can be done to minimize screen power consumption, beyond employing a screen saver and waiting for hardware improvements. Disk power consumption may be optimized by spinning down the disk whenever it has been inactive for several seconds; [2, 4, 5] have researched this topic. It is certainly possible to imagine ubiquitous computing devices with neither disks nor conventional displays; and, for such devices, minimizing the powerconsumption of the CPU will be particularly critical. Methods for saving CPU power have been suggested by [1, 3, ....

Kester Li, "Towards a low power file system," CS Tech Report 94-814, University of California, Berkeley, May 1994.

....mobile system. Forman et al. point out that the largest consumers in a typical laptop system are the disk drive and display [1] Li et al. show that spinning down the disk after it has been idle for as little as two seconds can have a dramatic effect on the power consumed by the disk drive [2] [3]. Douglis et al. also focus on reducing power consumption of disk drives, comparing the approach of spinning down after a fixed number of seconds with an optimal off line algorithm that minimizes latency and maximizes power savings, showing that the off line algorithm performs approximately 37 ....

Kester Li. Towards a low power file system. Technical Report UCB-CSD-94-914, University of California at Berkeley, May 1994.

....analogous) of the transfer, and the time of the access. This trace did not record deletions. The traces were preprocessed to convert file level accesses into disk level operations, by associating a unique disk location with each file. We used DOS traces collected by Kester Li at U.C. Berkeley [12], on IBM desktop PCs running Windows 3.1, also at file level. They include deletions. The traces were similarly preprocessed. We used disk level traces collected by Ruemmler and Wilkes on an HP workstation running HP UX [20] These traces include metadata operations, which the file level traces ....

....taking 20ms plus a value that is exponentially distributed with a mean of 3s. Though only the MAC trace comes from a mobile environment, the two desktop traces represent workloads similar to what would be used on mobile computers, and have been used in simulations of mobile computers in the past [12, 13, 15]. Table 3 lists additional statistics for the nonsynthetic traces. 4.2 Simulator Our simulator models a storage hierarchy containing a buffer cache and non volatile storage. The buffer cache is the first level searched on a read and is the target of all write operations. The cache is ....

Kester Li. Towards a low power file system. Technical Report UCB/CSD 94/814, University of California, Berkeley, CA, May 1994. Masters Thesis.

....file, the size of the transfer, and the time of the access. This trace did not record deletions. The traces were preprocessed to convert file level accesses into disk level operations, by associating a unique disk location with each file. We used dos traces collected by Kester Li at U.C. Berkeley [12], on IBM desktop PCs running Windows 3.1, also at file level. They include deletions. The traces were similarly preprocessed. We used disk level traces collected by Ruemmler and Wilkes on an hp workstation running HP UX [20] These traces include metadata operations, which the file level traces do ....

....taking 20ms plus a value that is exponentially distributed with a mean of 3s. Though only the mac trace comes from a mobile environment, the two desktop traces represent workloads similar to what would be used on mobile computers, and have been used in simulations of mobile computers in the past [12, 13, 15]. Table 3 lists additional statistics for the nonsynthetic traces. 4.2 Simulator Our simulator models a storage hierarchy containing a buffer cache and non volatile storage. The buffer cache is the first level searched on a read and is the target of all write operations. The cache is ....

Kester Li. Towards a low power file system. Technical Report UCB/CSD 94/814, University of California, Berkeley, CA, May 1994. Masters Thesis.

No context found.

Kester Li, Toward a low power file system, Master's thesis, University of California at Berkeley, May 1994. 22

No context found.

Kester Li, Toward a low power file system, Master's thesis, University of California at Berkeley, May 1994.

No context found.

K. Li, `Towards a low power file system', Technical Report UCB/CSD 94/814, University of California, Berkeley, CA, May 1994, Masters Thesis.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC