情報処理学会 第79回全国大会講演要旨

This study provides a comparative evaluation of job scheduling on two resource disaggregated datacenter infrastructures, i.e., a rackscale (RS) infrastructure and an inter-rackscale (IRS) infrastructure. The former emerges as a famous custom architecture where one rack consists of a number of processors, storages and accelerators that can be customized to a target application. The latter is further designed as an evolution of the former to disaggregate various hardware resources into different racks according to their own areas. For RS, its resource utilization is low because a whole rack is exclusively occupied by one job. For IRS, it uses free-space optics (FSO) for tightly-coupled connections between processors, storages and GPUs distributed in different racks, by swapping endpoints of FSO links to change network topologies. Through our evaluation by a large system simulation, we present the advantage of the FSO-equipped IRS architecture in job scheduling such as average queuing time of all dispatched jobs.