基于逐步细化快照序列的多核并行程序调试
发布时间：2018-08-25 09:33
【摘要】：多核并行程序的调试是一个公认的困难问题,困难主要来自于程序执行的不确定性.可重现调试(replay debug)提供了消除程序中不确定性的能力,但是现有的可重现调试解决方案都无法应用于商用的软硬件平台中,且进行调试所带来的性能损失会随着并发度的增加而超线性地增长.提出了一种基于运行快照的新型并行程序调试方法SDT(snapshot debug tool).该方法以离线的断点设置、运行快照捕捉和运行快照细化为基础,提出了一套可以指导用户由粗到细发现错误的调试过程,并在通用的软硬件平台上进行了实现.实验结果显示,在8线程的并发条件下,使用SDT调试所带来的时间性能损耗平均为51.88%;同时当线程数增长4倍时,使用SDT所带来的额外时间消耗最多增长1倍,具有很好的可扩展性.记录快照的数据量是影响SDT性能的重要挑战,实验证明通过使用增量式的快照记录方式可以有效地降低需要记录的数据量,减少记录快照花费的时间,提高SDT的整体性能.
[Abstract]:The debugging of multi-core parallel programs is a recognized difficulty, which comes mainly from the uncertainty of program execution. Reproducible debugging (replay debug) provides the ability to eliminate uncertainty in programs, but none of the existing reproducible debugging solutions can be applied to commercial software and hardware platforms. And the performance loss caused by debugging will increase superlinearly with the increase of concurrency. This paper presents a novel parallel program debugging method SDT (snapshot debug tool). Based on running snapshot. The method is based on off-line breakpoint setting, running snapshot capture and running snapshot refinement. A set of debugging processes that can guide users to detect errors from coarse to fine are proposed and implemented on a general software and hardware platform. The experimental results show that under the concurrent condition of 8 threads, the time performance loss caused by debugging with SDT is 51.88 on average, and when the number of threads increases four times, the extra time consumption caused by using SDT increases at most 1 times. It has good expansibility. Recording the amount of snapshot data is an important challenge to the performance of SDT. Experiments show that the incremental snapshot recording method can effectively reduce the amount of data needed to record, reduce the time spent recording snapshot, and improve the overall performance of SDT.
【作者单位】： 北京航空航天大学中德软件技术联合研究所;
【分类号】：TP311;;TP302;;TP306
[Abstract]:The debugging of multi-core parallel programs is a recognized difficulty, which comes mainly from the uncertainty of program execution. Reproducible debugging (replay debug) provides the ability to eliminate uncertainty in programs, but none of the existing reproducible debugging solutions can be applied to commercial software and hardware platforms. And the performance loss caused by debugging will increase superlinearly with the increase of concurrency. This paper presents a novel parallel program debugging method SDT (snapshot debug tool). Based on running snapshot. The method is based on off-line breakpoint setting, running snapshot capture and running snapshot refinement. A set of debugging processes that can guide users to detect errors from coarse to fine are proposed and implemented on a general software and hardware platform. The experimental results show that under the concurrent condition of 8 threads, the time performance loss caused by debugging with SDT is 51.88 on average, and when the number of threads increases four times, the extra time consumption caused by using SDT increases at most 1 times. It has good expansibility. Recording the amount of snapshot data is an important challenge to the performance of SDT. Experiments show that the incremental snapshot recording method can effectively reduce the amount of data needed to record, reduce the time spent recording snapshot, and improve the overall performance of SDT.
【作者单位】： 北京航空航天大学中德软件技术联合研究所;
【分类号】：TP311;;TP302;;TP306
【相似文献】
相关会议论文 前6条
1 王巍;方滨兴;张宏莉;;网格环境下的限定性并行程序可达性测试算法[A];全国网络与信息安全技术研讨会'2005论文集（上册）[C];2005年
2 王恩柱;刘晓平;;一个可视化并行程序开发平台的实现[A];计算机技术与应用进展——全国第17届计算机科学与技术应用（CACIS）学术会议论文集（上册）[C];2006年
3 郑启龙;王昊;吴晓伟;房明;;HPMR:多核集群上的高性能计算支撑平台[A];2008年全国开放式分布与并行计算机学术会议论文集(上册)[C];2008年
4 李正夫;郭权;王希诚;;新药研发网格下的一种药物分子对接并行程序[A];计算机技术与应用进展·2007——全国第18届计算机技术与应用（CACIS）学术会议论文集[C];2007年
5 赵鹏;张丹丹;汪鲁兵;田振夫;钱跃z，
本文编号：2202468
本文链接：https://www.wllwen.com/kejilunwen/jisuanjikexuelunwen/2202468.html
