More Like this

1. Avoiding Scheduler Subversion using Scheduler–Cooperative Locks

   Patel, Yuvraj ; Yang, Leon ; Arulraj, Leo ; Arpaci-Dusseau, Andrea ; Arpaci-Dusseau, Remzi ; Swift, Michael ( April 2020 , European Conference on Computer Systems (EuroSys))

   We introduce the scheduler subversion problem, where lock usage patterns determine which thread runs, thereby subverting CPU scheduling goals. To mitigate this problem, we introduce Scheduler-Cooperative Locks (SCLs), a new family of locking primitives that controls lock usage and thus aligns with system-wide scheduling goals; our initial work focuses on proportional share schedulers. Unlike existing locks, SCLs provide an equal (or proportional) time window called lock opportunity within which each thread can acquire the lock. We design and implement three different scheduler-cooperative locks that work well with proportional-share schedulers: a user-level mutex lock (u-SCL), a reader-writer lock (RWSCL), and a simplified kernel implementation (k-SCL). We demonstrate the effectiveness of SCLs in two user-space applications (UpScaleDB and KyotoCabinet) and the Linux kernel. In all three cases, regardless of lock usage patterns, SCLs ensure that each thread receives proportional lock allocations that match those of the CPU scheduler. Using microbenchmarks, we show that SCLs are efficient and achieve high performance with minimal overhead under extreme workloads. 

   more » « less
2. Avoiding Scheduler Subversion using Scheduler–Cooperative Locks

   Patel, Y. ; Yang, L. ; Arulraj, L. ; Arpaci-Dusseau, A. ; Arpaci-Dusseau, R. ; Swift, M. ( April 2020 , EuroSys)

   We introduce the scheduler subversion problem, where lock usage patterns determine which thread runs, thereby subverting CPU scheduling goals. To mitigate this problem, we introduce Scheduler-Cooperative Locks (SCLs), a new family of locking primitives that controls lock usage and thus aligns with system-wide scheduling goals; our initial work focuses on proportional share schedulers. Unlike existing locks, SCLs provide an equal (or proportional) time window called lock opportunity within which each thread can acquire the lock. We design and implement three different scheduler-cooperative locks that work well with proportional-share schedulers: a user-level mutex lock (u-SCL), a reader-writer lock (RWSCL), and a simplified kernel implementation (k-SCL). We demonstrate the effectiveness of SCLs in two user-space applications (UpScaleDB and KyotoCabinet) and the Linux kernel. In all three cases, regardless of lock usage patterns, SCLs ensure that each thread receives proportional lock allocations that match those of the CPU scheduler. Using microbenchmarks, we show that SCLs are efficient and achieve high performance with minimal overhead under extreme workloads. 

   more » « less
3. On Max-Min Fairness of Completion Times for Multi-Task Job Scheduling

   Shafiee, Mehrnoosh ; Ghaderi, Javad ( July 2020 , 2020 IFIP Networking Conference (Networking))

   We study the max-min fairness of multi-task jobs in distributed computing platforms. We consider a setting where each job consists of a set of parallel tasks that need to be processed on different servers, and the job is completed once all its tasks finish processing. Each job is associated with a utility which is a decreasing function of its completion time, and captures how sensitive it is to latency. The objective is to schedule tasks in a way that achieves max-min fairness for jobs' utilities, i.e., an optimal schedule in which any attempt to improve the utility of a job necessarily results in hurting the utility of some other job with smaller or equal utility. We first show a strong result regarding NP-hardness of finding the max-min fair vector of job utilities. The implication of this result is that achieving max-min fairness in many other distributed scheduling problems (e.g., coflow scheduling) is NP-hard. We then proceed to define two notions of approximation solutions: one based on finding a certain number of elements of the max-min fair vector, and the other based on a single-objective optimization whose solution gives the max-min fair vector. We develop scheduling algorithms that provide guarantees under these approximation notions, using dynamic programming and random perturbation of tasks' processing times. We verify the performance of our algorithms through extensive simulations, using a real traffic trace from a large Google cluster. 

   more » « less
4. On Max-Min Fairness of Completion Times for Multi-Task Job Scheduling

   Shafiee, Mehrnoosh ; Ghaderi, Javad ( July 2020 , IFIP Networking Conference)

   We study the max-min fairness of multi-task jobs in distributed computing platforms. We consider a setting where each job consists of a set of parallel tasks that need to be processed on different servers, and the job is completed once all its tasks finish processing. Each job is associated with a utility which is a decreasing function of its completion time, and captures how sensitive it is to latency. The objective is to schedule tasks in a way that achieves max-min fairness for jobs’ utilities, i.e., an optimal schedule in which any attempt to improve the utility of a job necessarily results in hurting the utility of some other job with smaller or equal utility.We first show a strong result regarding NP-hardness of finding the max-min fair vector of job utilities. The implication of this result is that achieving max-min fairness in many other distributed scheduling problems (e.g., coflow scheduling) is NP-hard. We then proceed to define two notions of approximation solutions: one based on finding a certain number of elements of the max-min fair vector, and the other based on a single-objective optimization whose solution gives the max-min fair vector. We develop scheduling algorithms that provide guarantees under these approximation notions, using dynamic programming and random perturbation of tasks’ processing times. We verify the performance of our algorithms through extensive simulations, using a real traffic trace from a large Google cluster. 

   more » « less
5. Themis: Fair and Efficient GPU Cluster Scheduling

   Mahajan, K ; Balasubramanian, A ; Singhvi, A ; Venkataraman, S ; Akella, A ; Phanishayee, A ; Chawla, S ( February 2020 , 17th USENIX Symposium on Networked Systems Design and Implementation)

   Modern distributed machine learning (ML) training workloads benefit significantly from leveraging GPUs. However, significant contention ensues when multiple such workloads are run atop a shared cluster of GPUs. A key question is how to fairly apportion GPUs across workloads. We find that established cluster scheduling disciplines are a poor fit because of ML workloads' unique attributes: ML jobs have long-running tasks that need to be gang-scheduled, and their performance is sensitive to tasks' relative placement. We propose Themis, a new scheduling framework for ML training workloads. It's GPU allocation policy enforces that ML workloads complete in a finish-time fair manner, a new notion we introduce. To capture placement sensitivity and ensure efficiency, Themis uses a two-level scheduling architecture where ML workloads bid on available resources that are offered in an auction run by a central arbiter. Our auction design allocates GPUs to winning bids by trading off fairness for efficiency in the short term, but ensuring finish-time fairness in the long term. Our evaluation on a production trace shows that Themis can improve fairness by more than 2.25X and is ~5% to 250% more cluster efficient in comparison to state-of-the-art schedulers. 

   more » « less