Trading Capacity for Performance in a Disk Array (OSDI'00)
Takeaway:
1. One could derive analytically models of disk performance (latency, throughput, etc) based on the disk and workload characteristics (rotational time, seek time, busyness of the system/queue size, read/write ratio, seek locality) and how we place data (striping, replication etc).
2. These models could be used to guide scheduling, and maybe accounting, by actually predicting what's going to happen when serving these requests.
3. Current head position, physical track layout etc., could be found by issuing requests to the disk.
Robust, Portable I/O Scheduling with the Disk Mimic (ATC'03)
by Remzi et. al
Takeaway:
1. Online disk service time prediction based on a more abstract model: instead of attempt to simulate the mechanism or components internal to the disk, simply reproduce the output as a function of the inputs observed.
2. Need to reduce the input space: most important inputs are just logical distance and request types.
3. This method accounts for most of the latency of the disk, but just in a probabilistic manner. And accuracy doesn't necessary translate to performance in scheduling.
4. As most paper from Remzi's group, it has a beautiful evaluation section where they really explored every possible configuration of their solution and understood what is going on.
Argon: performance insulation of shared storage servers (FAST'07)
from CMU
Takeaway:
1. The primary interference of storage system come from disk head time and cache space, which cannot be easily insulated.
2. The use prefetching/write-back to ensure a high disk head efficiency (which is kind of orthogonal to insulation), then cache partition and time slicing for insulation. With the parameters of the above three techniques dynamically adjusted by some analytic models.
3. They say that trade-off between efficiency and latency is fundamental.
4. Using feedback for scheduling is difficult here because you don't know what to expect due to the variability of performance from different workload, so a theoretical prediction could help here.
没有评论:
发表评论