Various techniques have been proposed in the literature to improve erasure code computation efficiency, including optimizing bitmatrix design and computation schedule, common XOR (exclusive-OR) operation reduction, caching management techniques, and vectorization techniques. These techniques were largely proposed individually, and, in this work, we seek to use them jointly. To accomplish this task, these techniques need to be thoroughly evaluated individually and their relation better understood. Building on extensive testing, we develop methods to systematically optimize the computation chain together with the underlying bitmatrix. This led to a simple design approach of optimizing the bitmatrix by minimizing a weighted computation cost function, and also a straightforward coding procedure—follow a computation schedule produced from the optimized bitmatrix to apply XOR-level vectorization. This procedure provides better performances than most existing techniques (e.g., those used in ISA-L and Jerasure libraries), and sometimes can even compete against well-known but less general codes such as EVENODD, RDP, and STAR codes. One particularly important observation is that vectorizing the XOR operations is a better choice than directly vectorizing finite field operations, not only because of the flexibility in choosing finite field size and the better encoding throughput, but also its minimal migration efforts onto newer CPUs.

We propose a delayed parity generation technique for maximum distance separable (MDS) storage codes, for two possible applications: the first is to improve the write-speed during data intake where only a subset of the parities are initially produced and stored into the system, and the rest can be produced from the stored data during a later time of lower system load; the second is to provide better adaptivity, where a lower number of parities can be chosen initially in a storage system, and more parities can be produced when the existing ones are not sufficient to guarantee the needed reliability or performance. In both applications, it is important to reduce the data access as much as possible during the delayed parity generation procedure. For this purpose, we first identify the fundamental limit for delayed parity generation through a connection to the well-known multicast network coding problem, then provide an explicit and low-complexity code transformation that is applicable on any MDS codes to obtain optimal codes. The problem we consider is closely related to the regenerating code problem, however the proposed codes are much simpler and have a much smaller subpacketization factor than regenerating codes, and thus our result in fact shows that blindly adopting regenerating codes in these two settings is unnecessary and wasteful.