Balancing HPC applications through smart allocation of resources in MT processors

Abstract

Abstract-Many studies have shown that load imbalancing causes significant performance degradation in High Performance Computing (HPC) applications. Nowadays, Multi-Threaded (MT 1 ) processors are widely used in HPC for their good performance/energy consumption and performance/cost ratios achieved sharing internal resources, like the instruction window or the physical register. Some of these processors provide the software hardware mechanisms for controlling the allocation of processor's internal resources. In this paper, we show, for the first time, that by appropriately using these mechanisms, we are able to control the tasks speed, reducing the imbalance in parallel applications transparently to the user and, hence, reducing the total execution time. Our results show that our proposal leads to a performance improvement up to 18% for one of the NAS benchmark. For a real HPC application (much more dynamic than the benchmark) the performance improvement is 8.1%. Our results also show that, if resource allocation is not used properly, the imbalance of applications is worsened causing performance loss.

Enhancing the performance of assisted execution runtime systems through hardware/software techniques

“…STM 2 and all the applications are compiled with GCC 4.3.4 with optimization level O3; the results reported for each application are the average of 25 runs. In order to use all hardware priority levels, all tests are performed on a custom version of the Linux 2.6.33 kernel patched with the HMT patch [3]. In all the experiments, Eigenbench and STAMP applications use all the 32 available hardware threads: 16 application threads and 16 auxiliary threads.…”

“…As Table 1 shows, not all hardware thread priority values can be set by applications: user software can only set priority levels 2, 3, 4; the operating system (OS) can set 6 out of 8 levels, from 1 to 6; the Hypervisor can span the whole range of priorities. In order to use all possible levels of priorities, a special Linux 2.6.33 kernel patched with the Hardware Managed Threads priority (HMT) patch [2,3,4] is required. This custom kernel provides two interfaces (a sysfs and a system call) through which the users can set the current hardware thread priority, including the ones that require OS or Hypervisor privilege (the OS issues a special Hypervisor call to set priority 0 and 7).…”

“…Other researchers [23] have also investigated the effect of hardware thread priorities on the execution time of co-scheduled application pairs on a trace-driven simulator of the POWER5 processor. Moreover, in a follow-up work, Boneti et al used hardware prioritization to transparently balance high performance computing applications [3,4], achieving up to 18% performance improvement.…”

Load balancing using dynamic cache allocation

“…According to the classification in [3], there are two main classes of load imbalance due to causes that become apparent only during the application's execution: (i) Intrinsic load imbalance is caused by characteristics, which are intrinsic to the application, such as the input data. For example, sparse matrix computations heavily depend on the number of non-zero values in the matrix; the convergence time of iterative methods that approximate the solution of a problem may change for different domains of the modeled space.…”

“…For example, the OS may decide to run another process (say a kernel daemon) in place of the process running on a CPU. Also, extrinsic load imbalance may be caused by thread contention for processor's shared resources; this may be particularly true in the case of SMT architectures, where threads share and compete for most of the processor's resources [3]. Clearly, there is nothing that the application programmer could do a priori to prevent extrinsic load imbalance.…”