Predicting Slow Network Transfers in Scientific Computing

Abstract

Data access throughput is one of the key performance metrics in scientific computing, particularly for distributed data-intensive applications. While there has been a body of studies focusing on elephant connections that consume a significant fraction of network bandwidth, this study focuses on predicting slow connections that create bottlenecks in distributed workflows. In this study, we analyze network traffic logs collected between January 2019 and May 2021 at National Energy Research Scientific Computing Center (NERSC). Based on the observed patterns from this data collection, we define a set of features to be used for identifying low-performing data transfers. Through extensive feature engineering and feature selection, we identify a number of new features to significantly enhance the prediction performance. With these new features, even the relatively simple decision tree model could predict slow connections with a F1 score as high as 0.945.

Leveraging History to Predict Infrequent Abnormal Transfers in Distributed Workflows

“…To evaluate their effectiveness, we trained with all combinations of training and testing sets shown in Table 2 . We primarily utilized the try-all decision tree method, which previously demonstrated the best performance during the first phase of our research [ 12 ].…”

“…For example, we can declare all transfers whose throughput is less than 1 × 10 6 bps as slow and all other transfers as non-slow (“normal”). This is the choice used in our initial work [ 12 ]. We will continue to use this choice but also attempt to explore alternative options as described next.…”

“…In our previous study with data transfer involving DTN01, we found most of the slow transfers are from IP addresses that are infrequently used, often appearing only once or a small number of times [ 12 ]. These occasional uses might involve a personal laptop in a work-from-home scenario or a user in an internet cafe.…”

“…The traffic logs are from dedicated data transfer nodes (DTNs) for moving the large amounts of data needed for mission-critical scientific applications [ 14 ]. This study expands on the workshop report [ 12 ] as follows: (1) The previous work only used data records from one of the DTNs at NERSC (dtn01) and only covered five months, from January–May 2021, while this study is based on an extensive dataset collected from all four public DTNs covering nearly two-year’s time (from January 2021 to August 2022). (2) This work systematically addresses the issue of far fewer slow transfers than normal ones by defining a set of sampling strategies for mitigating the class-imbalance concern.…”

“… We devise a strategy to capture the network state by utilizing information from recently completed file transfers. (Even though this was a key technique used previously in [ 12 ], we feel it is still worthwhile to draw attention to it because it is an effective approach that could be effectively used in many application scenarios.) We define a set of features engineered from the most recent transfer from the same subnet as the transfers from a similar location, and the host may show similar behaviors.…”