Terms

../_images/workload-terms.png

Fig. 1 A sample write workload to illustrate the terms. This workload consists of two rounds. Each round has a work amount of 500 MB.

We define the following terms.

  • Performance index (PI) is one property we want to measure. For instance, the throughput of a storage device is a PI, and its latency is another PI.

Note

When multiple PIs are involved, Pilot applies the analytical methods you choose to each PI and combines the results. For instance, when calculating the work amount for next round of benchmark, Pilot calculates the desired work amount for each PI and the largest one would be picked.

  • Session is the context for doing one measurement. We can measure multiple PIs in one session. One session can include multiple rounds of benchmarks, and each round can have a different length.
  • Work amount is the amount of work involved in one round of benchmark. The work amount is related to the length of the workload. For instance, in a sequential write I/O workload round, we write 500 MB data using 1 MB writes, the work amount of this round is 500.
  • Work unit is a smallest unit of work amount that we can get a measurement from. Using the above sample, if the I/O size is 1 MB, we can measure the time of each I/O syscall and calculate the throughput of each I/O operation. Here the work unit is 1 MB, and we have 500 work units in that workload round. Not all workloads should be divided into units. Pilot expects the work unit to be reasonably homogeneous. So, for instance, reading one 1 MB from different locations of a device can be thought as homogeneous because the difference in performance is small and mostly normally distributed. But shifting from sequential I/O to random is not homogeneous because that would result in huge difference in I/O performance. In general, the user should only divide the workload into units when one emph{expects} them to have similar performance. If not the user should not use the work unit-related analytical methods of Pilot and should stick with readings (see next term) only. We leave heterogeneous work units as a future work.
  • Reading is a measurement of a PI of a round. Each benchmark round generates one reading for each PI at the end of the round. In the sample above, when PI is the throughput of the device, we can get one throughput reading for each round.
  • Unit reading (UR) is a measurement of a PI of a work unit. In the sample above, we would have 500 throughput Unit readings for Round 1 because it contains 500 work units, and these 500 unit readings would be the throughput of each 1 MB I/O operation.
  • Work-per-second (WPS) is the calculated speed at which the workload consumes work. This is usually the desired PI for many simple workloads.

Some workload cannot be meaningfully separated into homogeneous work units, such as booting up a system and randomly reading files of different sizes. We get only readings from this kind of workloads.