Mapping and scheduling HPC applications for optimizing I/O

Abstract

In HPC platforms, concurrent applications are sharing the same file system. This can lead to conflicts, especially as applications are more and more data intensive. I/O contention can represent a performance bottleneck. The access to bandwidth can be split in two complementary yet distinct problems. The mapping problem and the scheduling problem. The mapping problem consists in selecting the set of applications that are in competition for the I/O resource. The scheduling problem consists then, given I/O requests on the same resource, in determining the order to these accesses to minimize the I/O time. In this work we propose to couple a novel bandwidth-aware mapping algorithm to I/O list-scheduling policies to develop a cross-layer optimization solution.We study this solution experimentally using an I/O middleware: CLARISSE. We show that naive policies such as FIFO perform relatively well in order to schedule I/O movements, and that the important part to reduce congestion lies mostly on the mapping part. We evaluate the algorithm that we propose using a simulator that we validated experimentally. This evaluation shows important gains for the simple, bandwidth-aware mapping solution that we provide compared to its non bandwidth-aware counterpart. The gains are both in terms of machine efficiency (makespan) and application efficiency (stretch). This stresses even more the importance of designing efficient, bandwidth-aware mapping strategies to alleviate the cost of I/O congestion.

Drishti: Guiding End-Users in the I/O Optimization Journey

“…Too many I/O phases [9] ✓ ✓ Bandwidth limited by a single OST I/O bandwidth [9], [10] ✓ Limited by the small data size [9] ✓ ✓ Unbalanced I/O workload among ranks [9] ✓ ✓ Large number of small I/O requests [9] ✓ ✓ Unbalanced I/O workload on OSTs [9], [11] ✓ ✓ ✓ Bad file system weather [9], [12] ✓ Redundant/overlapping I/O accesses [24], [25] ✓ ✓ I/O resource contention at OSTs [13], [14] ✓ Heavy metadata load [10] ✓…”

Protecting Metadata Servers From Harm Through Application-level I/O Control

“…PADLL is able to control the rate of both data and metadata workflows. Other systems are directly implemented within core layers of the HPC I/O stack, including the PFS [14], [18], [20], [22], [23], scheduler [21], and I/O libraries [16], [17]. These solutions are intrusive and offer limited maintainability and portability.…”

“…If usage i is higher than demand i , the controller assigns the minimum between demand i and the fair_share (8-9). The algorithm then distributes leftover rate (left R ) across actives jobs (11)(12)(13)(14). Specifically, it computes the overall rate used by all jobs (11), and assigns left R based on their usage proportion, given by usage_proportion i (13)(14).…”

“…The algorithm then distributes leftover rate (left R ) across actives jobs (11)(12)(13)(14). Specifically, it computes the overall rate used by all jobs (11), and assigns left R based on their usage proportion, given by usage_proportion i (13)(14). Finally, the global controller generates rules (enforce) to be submitted to each local controller (15), and sleeps for loop_interval before beginning a new control cycle (16).…”

“…Metadata remains overlooked. While existing proposals focus on achieving QoS over data workflows (I/O bandwidth) [14,21,23,29,43,44,53,54], the metadata counterpart has not received the same level of attention. In fact, several HPC centers are observing a surge of metadata operations in their clusters and expect this to become more severe over time.…”