Xiao Qin's Research

Auburn University

Final Report

BUD: A Buffer-Disk Architecture for Energy Conservation in Parallel Disk Systems

Download BUD Final Report PDF




Findings

An Energy-Efficient Framework for Large-Scale Parallel Storage Systems
A Simulation Framework for Energy-efficient Data Grids
An Energy-Efficient Scheduling Algorithm Using Dynamic Voltage Scaling for Parallel Applications on Clusters
Load-Balancing Strategies for Energy-Efficient Parallel Storage Systems with Buffer Disks
Sacrificing Reliability for Energy Saving: Is It Worthwhile for Disk Arrays
Load-Balancing Strategies for Energy-Efficient Parallel Storage Systems with Buffer Disks
Energy-Efficient Prefetching for Parallel I/O Systems with Buffer Disks
Improving Reliability and Energy Efficiency of Disk Systems via Utilization Control
Energy Conservation for Real-Time Disk Systems with I/O Burstiness
An Adaptive Energy-Conserving Strategy for Parallel Disk Systems
DARAW: A New Write Buffer to Improve Parallel I/O Energy-Efficiency
MICRO: A Multi-level Caching-based Reconstruction Optimization for Mobile Storage Systems
PEARL: Performance, Energy, and Reliability Balanced Dynamic Data Redistribution for Next Generation Disk Arrays
HyBUD: An Energy-Efficient Architecture for Hybrid Parallel Disk Systems
Energy-Aware Prefetching for Parallel Disk Systems
DORA: A Dynamic File Assignment Strategy with Replication
Collaboration-Oriented Data Recovery for Mobile Disk Arrays
A File Assignment Strategy Independent of Workload Characteristic Assumptions
References


[Manzanares et al., 2008a] The prefetching algorithm uses the frequency that a block is requested as a heuristic. The first step of the algorithm iterates over all of the requests and counts the references for each unique block. Then it sorts the list of unique blocks by the number of references to each block. At this point the algorithm puts the highly requested blocks into the buffer until it is full.

The last part of the algorithm also iterates over all requests trying to figure out how long each disk can sleep. If a block requested for a disk is in the buffer disk, the corresponding data disk can sleep longer. The buffer disk handles the request and the distance between requests on the same disk becomes cumulative. If a requested block is not in the buffer the disk must be woken up to serve the request, this is handled by the power management mechanism. The distance is then set to zero since the disk had to be woken up. Using the frequently accessed heuristic the PRE-BUD strategy should have a small performance impact on the system. Almost all steps of the algorithm run linearly with respect to the number of requests. The only step that is not linear is the phase that sorts the list of requests according to their frequency. Sorting is a common procedure and is known to have a best-case run-time of O(nlogn), where n is the number of data request to be sorted. The PRE-BUD strategy is able to have a run time of O(n + nlogn) using an efficient sorting algorithm. The PRE-BUD strategy is not assumed to be optimal, since the requested blocks are sorted using their frequency. The frequency is used as a heuristic to select blocks to be placed in the buffer. An optimal strategies goal would be to select the requests to be placed in the buffer that produce the largest impact on the standby time of disks. The standby time of each disk is directly related to the energy savings of the system.

The results displayed in the Figure below held the number of disks to 4 and also kept the data size of each request at 275MB. We have omitted the results of the non-energy aware approach, since they are constant and higher than the energy aware strategy. Buffer Strategy 1 adds an extra disk and Buffer Strategy 2 uses an existing disk as the buffer disk. As expected the performance of both of our strategies were lowered when the hit rate was decreased. Disk sleep times are lowered once a miss is encountered. This is due to the fact that a disk has to wake up to serve the request. This will increase the energy consumption of disks that have to serve the missed requests. This leads to an increase in the total energy consumption of the entire system. Our buffered large-scale parallel disk system is still able to consume less energy than the energy aware approach. The energy aware and non-energy aware disk systems are not affected by buffer disk miss rates.

As the hit rate is lowered, the first buffer disk approach begins to approach the same level of performance as the energy aware strategy. It is only able to save 10% energy over the energy aware strategy when the hit rate is 75%. This is because adding the extra disk puts extra energy requirements on the system, and lowering the hit rate further impacts the energy benefits of the first strategy. The second buffered disk approach is still able perform 25% better than the energy aware approach. This is because there is not the extra energy penalty of adding an extra disk. The capacity of your disk system will be lowered using this approach.

The hit rate becomes a very important factor in the performance of our approaches. If the buffer disk is constantly missing requests then both strategies will eventually downgrade to the energy aware approach. Fortunately applications have been documented to request 20% of the data available 80% of the time. Our heuristic based approach would work considerably well in this case. This is modeled by the 80% hit rate. The buffered disk approach one and two are able to save 12% and 26% energy over the energy aware strategy when the hit rate is 80%. Similarly, they are able to save 37% and 47% of the total energy compared to the non-energy aware approach.

The first buffer disk approach downgrades more quickly than the second approach as the hit rate is decreased as compared to the energy aware approach. This is not that great of a concern since the first strategy is still able to have a positive impact on the reliability of the system as compared to the energy aware approach. The first buffer disk approach is still able to produce significant energy savings over the non-energy aware approach without compromising the reliability of the system. The energy savings performance of the second approach does not diminish as quickly as the first approach, but there will be an impact on the capacity of the system. The second approach is also able to reduce the number of state transitions.



From the above Figure we are able to see that the non-energy aware approach wastes a considerably larger amount of energy as compared to all of the energy aware approaches. This is expected since the non-energy aware approach is not able to place disks in the standby mode. Buffer strategy 1 was able to produce a 12% increase in energy savings over the energy aware strategy when 10 disks were simulated. Similarly, buffer strategy performed even better with an 18% increase. This is expected again because of the energy overhead adding an extra disk Buffer Strategy 1 requires. Our approach produces promising results as the number of disks is increased. This is an important observation, since our target system is a large-scale parallel system. This leads us to believe our system will produce energy benefits regardless of the number of disks in a system.

[Bellam et al., 2008] RAID 1 is popular and is widely used for disk drives. RAID 1 is implemented with a minimum of two disks, which are the primary and back disks. Initially the data is stored to the primary disk and then it is mirrored to the backup disk. This mirroring helps to recover the data when there is a failure in the primary disk. It also helps to increase the performance of the RAID 1 system by sharing the workload between the disks. We considered RAID 1 for all of our experiments.
The processor in the system generates the I/O stream, which is queued to the buffer. The utilization of the disk is calculated using the request arrival rate. Please refer to Section 3 for details of the description for the disk utilization model.
It should be noted that all requests here are considered as read requests. At any given point of time the disks can be in the following three states.
• State 1: Both the disks in sleep mode
• State 2: Primary disk active and backup disk in sleep mode
• State 3: Both the disks in active mode and share the load.
Let us consider that the disks are in state 1 at the beginning. Once the utilization is calculated, it is compared with the safe utilization zone range. If the calculated value falls below the range then disks stay in state 1. If the calculated value is within the range, then the primary disk is made active while the backup disk continues to stay in the sleep mode. This represents a transition to state 2. If the calculated value is beyond the range then both the disks are made active and both of them share the load, which corresponds to state 3. Transition of states from one power mode to another involves disk spin up and/or spin down. The disk spin ups and spin downs also consume a lot of energy.

RAID 1 is used in our experiments. RAID 1 uses a minimum of two disks, one as a primary and one as the backup. We conducted the experiments on three types of disks from IBM. The experimental results are compared against two traditional state of the art methods. In the first method, load balancing, both the disks are always made active. Load balancing achieves very high performance because both the disks share the load. The second method, traditional method, is where the primary disk is made always active and the backup disk is kept in sleep mode. The backup disk is made active only when the utilization of the primary disk exceeds 100%, also known as saturation. In what follows, we term our approach as RAREE (Reliability aware energy efficient approach).




The experimental data generated from the simulations is plotted in the above Figure, which represents the energy consumed by the 2 year old disk respectively. RAREE is compared against load balancing and the traditional method. From the above Figure it is observed that for the IBM 36Z15 disk the power consumed by RAREE falls in between the load balancing and traditional techniques. Even for the IBM 73LZX the trend is similar, but the difference in values is not as high as IBM 36Z15. For the IBM 40GNX the power consumed by RAREE is smaller than the traditional and load balancing power consumption values because disk spin up and spin down values are much smaller for the IBM 40GNX when compared with the other two disks. It should be observed that the disk spin down and disk spin up values play a vital role in the energy consumption.




The above Figure shows the performance of RAREE on Travelstar disks of different ages. It can be observed from figure that RAREE consumes less energy when compared to traditional and load balancing.



The Figure above shows the effects on energy when the spin up energy is varied for a 2 year old disk. The RAREE energy consumption falls in between traditional and load balancing techniques. Though the energy consumed by RAREE is a little higher than traditional technique, here we are also gaining good amount reliability.



The above Figure shows the change in energy as the active power is varied. Here RAREE energy consumption is definitely less than the two existing techniques, because RAREE makes the disks go to sleep mode as soon there are no requests unlike the other techniques. When idle power is changed unlike the active power the energy consumed by the RAREE again falls between the two techniques. This is because RAREE makes the system go to sleep mode very often depending on the conditions. It should be observed here though the RAREE consumed a bit higher energy than traditional it can be neglected as we are achieving reliability.



The above Figure is a very important graph here it shows the reliability in terms of annual failure rate percentile. It can be observed from the graph that RAREE achieves a very high reliability when compared to load balancing and traditional. Only one bar is shown for load balancing and traditional techniques because both have the same reliability levels as they don’t pay special attention to reliability. We also found an interesting observation that when RAREE is applied to IBM40GNX, which is a travelstar, it definitely consumes much less energy than the other two Ultrastars, which are high performance disks. This makes it clear that RAREE gives best results when it is used on mobile disks instead of high performance disks. This doesn’t limit the usage of RAREE to mobile disks because though Ultrastar consumes a little more energy than traditional technique we still get a good reliability at a marginal cost of energy. Simulation results prove that on an average roughly 20% of energy can be saved when RAREe is used instead of load balancing. When RAREe is used instead of the traditional method an excess of 3% of energy is saved, it is not a very significant amount but along with a very little energy saving we are also achieving high reliability which makes it significant.