Next: Details
of the Polus Up:
Polus : Growing Storage Previous: Paper Organization
In this paper, the effectiveness of the competing ECA and Polus management approaches will be discussed using the example of a storage area network (SAN) file system [12,21]. As shown in figure 3, in a SAN file system, the clients contact a metadata server to obtain the necessary metadata for a particular file. Subsequently the clients go directly to the storage controllers via a SAN protocol to access the storage. The clients cache both the file metadata information and the user block data in two separate caches. In order to write ECA rules for this system, a system administrator needs to do the following:
Establish Goals: System administrators are usually interested in ensuring that certain performance (throughput, latency), reliability and security goals are being met in their SAN file system deployments. For example, they could specify their QoS goals as: (1) ensure that each client has a throughput of at least 40MBps; and (2) ensure that the system has 99.999 percent availability.
Determine the observables to analyze: System administrators have access to many static and dynamic system observables such as the available memory size, the SAN bandwidth being provided to a particular client, and the cache hit rate at the client, the metadata server or the storage controller. The system administrators also have to access to workload characteristics such as the read/write ratio, workload type (random or sequential), and the block size.
Assess the available actions: System administrators need to be aware of the different possible storage actions that they can perform to manage the storage, such as replication, migration, clean delay [18], request throttling, zoning etc.
Determine thresholds for the observables:
Based on prior empirical data or experience, system administrators need to
determine the threshold values which, when violated should result in the
triggering of corrective management actions. For example, if then take a corrective
action.
Select a particular action: If the threshold value of a particular observable is being violated then the system administrator needs to choose a correction action such as increasing the prefetch size or replication of data.
Determine the granularity of the action: For example, when the corrective action being taken is to increase the data prefetch size, then the system administrator needs to also specify the unit of the prefetch size increase.
To put it all together, if the QoS goal of 10 millisecond latency is not being met for a particular client, then the system administrator needs to write the following sets of ECA rules (not exhaustive) to find a remedy:
[Rule 1] If the throughput of a storage controller is at its maximum, then migrate this client's data to another controller that has the necessary available bandwidth.
[Rule 2] If the and the workload
is sequential then increase the data prefetch size by
4 objects.
[Rule 3] If a particular client is exceeding its allotted bandwidth then throttle its request.
Thus, for a particular QoS goal, the system administrator needs to evaluate the values of all the relevant system observables, assess whether they are violating a particular predetermined threshold value, and then choose a corrective action from a list of possible system management actions. The objective of Polus is to reduce the number of details that a system administrator needs to consider.
Next: Details of the Polus Up: Polus :
Growing Storage Previous: Paper Organization