The Co-Design of Exascale Storage Architectures (CODES) simulation framework builds upon the ROSS parallel discrete event simulation engine to provide high-performance simulation utilities and models for building scalable distributed systems simulations
We say to use fio, but don't have more specifics than that.
Some notes from Kevin:
“We measured I/O performance for 60 seconds at each access size as the access size was varied from 4 KiB to 8 MiB for each of sequential read, sequen- tial write, random read, and random write access patterns. The libaio engine was used in all cases with an iodepth of 4 and direct I/O. We set the model’s rate parameter to the maximum rate achieved by the sequential read and sequential write measurements. The overhead parameter was set to the median completion latency minus the transfer rate for the 4K request size. The seek parameters were set to the difference between the sequential and random latency values divided by the iodepth (4) at each request size.”
In fio, one of the output timings is “clat” or completion latency. So you take the optimal read or write transfer rate and compute a time to transfer 4K based on that, then subtract it from the clat value. That became our overhead parameter.
Original Issue Author: Jonathan Jenkins Original Issue ID: 146 Original Issue URL: https://xgitlab.cels.anl.gov/codes/codes/issues/146
We say to use fio, but don't have more specifics than that.
Some notes from Kevin:
“We measured I/O performance for 60 seconds at each access size as the access size was varied from 4 KiB to 8 MiB for each of sequential read, sequen- tial write, random read, and random write access patterns. The libaio engine was used in all cases with an iodepth of 4 and direct I/O. We set the model’s rate parameter to the maximum rate achieved by the sequential read and sequential write measurements. The overhead parameter was set to the median completion latency minus the transfer rate for the 4K request size. The seek parameters were set to the difference between the sequential and random latency values divided by the iodepth (4) at each request size.”
In fio, one of the output timings is “clat” or completion latency. So you take the optimal read or write transfer rate and compute a time to transfer 4K based on that, then subtract it from the clat value. That became our overhead parameter.
Overhead = clat - (4K/best_transfer_rate)