MinIO进行Memory层面的性能评估:我们发现,对于特定于DNN的应用程序级缓存(如MinIO[39]),当我们在固定的CPU分配下改变分配给作业的内存量时,很容易对作业吞吐量行为进行建模。这是因为,MinIO确保了一个作业在每个时期获得固定数量的缓存命中。Synergy有意识地决定使用应用程序级MinIO缓存而不是页面缓存,因为MinIO在共享机器的独立作业之间提供内存隔离。如果我们不使用MinIO,我们将不得不在离散内存分配时对模型进行分析,这可能导致分析成本增加,并可能改变分析矩阵的趋势。然而,在Synergy中使用MinIO可以使缓存性能可预测,从而降低Synergy's的分析成本——允许乐观的分析。We observe that, with DNN-specific, application-level caches like MinIO [39], it is easy to model the job throughput behaviour as we vary the amount of memory allocated to a job at fixed CPU allocation. This is because, MinIO ensures that a job gets a fixed number of cache hits per epoch. Synergy makes a conscious decision to use application-level MinIO cache instead of Page Cache because MinIO provides memory isolation across independent jobs sharing the machine. If we do not use MinIO, we will have to profile the model at discrete memory allocations which could result in increased profiling costs, and also potentially change the trends in profiling matrix. However, the use of MinIO in Synergy makes cache performance predictable and hence reduces Synergy ’s profiling costs – allowing optimistic profiling.