Abstract In large-scale deep learning applications, SSDs (Solid State Drives) have been widely adopted to speed up the training. However, a snapshot process is periodically performed in a deep learning application, by which a great number of training parameters (at the TB level) need to be written to SSDs; thus, it poses serious challenges for the lifetime of SSDs. In this paper, we for the first time design a mechanism, called delta snapshot, that can effectively reduce the overall amount of data written to an SSD during the training phase so as to preserve its lifetime. Delta snapshot exploits the redundant information between snapshots. Specifically, we observe that the exponent part and the most significant bits of the mantissa change very little between two consecutive snapshots. Based on this, we develop effective mechanism to compress the redundant bits of snapshots to reduce the size of the written data. Experimental results showed that our technique can reduce the overall amount of written data by 31% and the erase operations by 27%, with a negligible time overhead in the training phase.
Paper
Full text
Preserving SSD lifetime in deep learning applications with delta snapshots
Semantic Scholar · Computer Science · 2019
Abstract
Abstract In large-scale deep learning applications, SSDs (Solid State Drives) have been widely adopted to speed up the training. However, a snapshot process is periodically performed in a deep learning application, by which a great number of training parameters (at the TB level) need to be written to SSDs; thus, it poses serious challenges for the lifetime of SSDs. In this paper, we for the first time design a mechanism, called delta snapshot, that can effectively reduce the overall amount of data written to an SSD during the training phase so as to preserve its lifetime. Delta snapshot exploits the redundant information between snapshots. Specifically, we observe that the exponent part and the most significant bits of the mantissa change very little between two consecutive snapshots. Based on this, we develop effective mechanism to compress the redundant bits of snapshots to reduce the size of the written data. Experimental results showed that our technique can reduce the overall amount of written data by 31% and the erase operations by 27%, with a negligible time overhead in the training phase.