Conserving Disk Energy in Network Servers ACM 17th annual international conference on Supercomputing Presented by Hsu Hao Chen

Outline Introduction Conserving disk energy Idle Replace Combined Multi-speed Evaluation Combined Multi-speed Conclusions

Introduction(1/2) The Google search engine 15K servers These large clusters consume a significant amount of energy Energy costs can reach 60% of the operational cost of data center Evaluating four approaches to solving it Idle Replace Combined Multi-speed

Introduction(2/2) Multi-speed Combined

Idle Most of them are based on powering disks down during periods of idleness Break-even threshold cost of powering the disk down and up (on the next access) Testing Assuming that load peaks reach only 50% of this maximum throughput A large memory cache Memory cache miss rate that is lower than 0.03% Result Average idle time of 15.2 seconds In summary, not appropriate for network servers

Replace(1/2) Replace each high-performance disk with one or more lower power disks 1-to-1 ratio for laptop disk problem Storage capacity Performance: access latency 1-to-2 ratio for low power SCSI disk Energy consuming Two low power disks > high-performance disk

Replace(2/2) 1-to-n ratio for laptop disk reliability problem We would need at least four (RAID) laptop disks for each HP

Combined(1/4) The idea is to associate each high- performance disk with a lower power disk, called a secondary disk. The disks should have the same size and mirror each other Coherence actions updates while the set of disks coming up was powered off

Combined(2/4) Implementation

Combined(3/4) Linux module Allows the creation of multiple virtual devices Each virtual device is mapped to a pair of disks Module is inserted at a low level, all disk traffic (including metadata accesses) is visible to it Module intercepts all calls to the ll_rw_block() kernel routine

Combined(4/4) The module has three key components A translation table per virtual device specifies which physical disk drive to use on each access ll_rw_block() Monitors the load on the disks and measure of the load offered EWMA α=0.875 Selects which disk to use depending on the load on the disk subsystem A bitmap per disk specifying all the blocks that have been written since the disks of the corresponding virtual device were last made coherent. A bit is set in a bitmap when an intercepted ll_rw_block() call produces a disk write

Multi-speed This approach does not require mutiple disks, coherence, bitmap etc.. Switching threshold Decides to changes speeds Emulation Because multi-speed disks are not available in the market The emulation keeps our two SCSI disks powered on all the time All write accesses are immediately directed to both disks The emulation also assigns performance and energy costs to the speed transitions

Evaluation(1/7) Network server hardware P4-1.9GHz 512MB RAM OS: Linux Storage disks the SCSI Ultrastar 36Z15 disk the SCSI Ultrastar 73LZX disk (when evaluating Multi-speed) the laptop Travelstar 40GNX disk (when evaluating Combined) a Gigabit Ethernet network interface Web server is able to service a maximum of 2520 requests/second for the Clarknet trace ClarkNet:These two traces contain two week's worth of all HTTP requests to the ClarkNet WWW server Proxy server can service up to 335 requests/second for the Hummingbird trace

Evaluation(2/7) Combined Power-saving 1%

Evaluation(3/7) Combined In summary, we do not consider very realistic for network servers. Power-saving 41%

Evaluation(4/7) Multi-speed for web-server Power-saving 16%Power-saving 22%

Evaluation(5/7) Multi-speed for proxy-server Power-saving 15% Power-saving 17%

Evaluation(6/7) Multi-speed web-serverproxy-server Throughput Degradation 1%Throughput Degradation 3%

Evaluation(7/7) Multi-speed In summary, the two-speed disk should perform well in a wide range of scenarios.

Conclusions Two-speed disk techniques energy savings without performance degradation in network servers Our results suggest that this technique should be carefully considered by disk manufacturers The other techniques we studied cannot provide any disk energy savings