[Letter from David P. Appleby]

Scalable Incremental Checkpointing using GPU-Accelerated De-Duplication

“…Efficient hash calculation using GPUs: We use the 128-bit Murmur3 [1] hash function for comparing chunks. Murmur3 is a common non-cryptographic hash function used for hash tables.…”

“…Over time, checkpointing has found broad applicability in other scenarios: batch job preemption [7] (e.g., to make room for higher priority on-demand jobs without losing computational progress), job migration [27], adjoint computations and automated differentiation methods that generate intermediate states during a forward pass and revisit them in a backward pass [35], exploring alternative computational paths (e.g., sensitivity analysis of AI models using variations of training data starting from a common initial training) [28], and the study of reproducibility by capturing and comparing intermediate results during different runs. Under such circumstances, checkpointing is more challenging for two reasons: (1) there is a need to store the entire checkpoint record into a lineage [18] (not just the latest checkpoint); and (2) the checkpointing frequency is significantly higher than in the case of resilience (e.g., checkpoint intervals of 10ms are common in adjoint computations [13] and reproducibility, as opposed to resilience, where checkpoint intervals are correlated with the mean time between failures and are the order of hours).…”

“…Even in the case of resilience, where the checkpointing frequency is low, I/O overheads are significant enough to warrant asynchronous multi-level checkpointing methods [21]: the processes write the checkpoints to the fastest local memory (e.g., GPU memory), then let the application continue running, while in the background they flush the checkpoints asynchronously to slower memory tiers and eventually the external repository. However, at high checkpointing frequency, such methods have two limitations: (1) there is only a limited amount of spare space available on the fastest memory tiers to cache checkpoints, so the HPC workflow may be delayed if it produces new checkpoints faster than they can be flushed to slower memory tiers; and (2) since the entire checkpoint record needs to be persisted, its accumulated size may quickly explode to large sizes and produce unacceptable resource utilization, even if performance considerations were not a concern. Thus, there is a need to simultaneously reduce both the I/O overheads and space utilization of checkpoints.…”

Framing Political Violence: Success and Failure of Religious Mobilization in the Philippines and Thailand

“…Additionally, religion might help to unite otherwise fragmented groups by putting the faithful against a common enemy who is portrayed as a force of evil determined to destroy a sacred community. 29 In a second step, appropriate reactions to the identified problems are articulated. Peaceful solutions should be portrayed as being hopeless, thereby highlighting the need for violent action.…”

Advanced Membrane Materials for Fuel Cell Applications

“…The performance evaluation carried out in experimental size Ballard MK4 fuel cell with an active area of 50 cm 2 operated on air/hydrogen at 24/42 psig and at 70˚C showed that BAM1G membranes of 39-420 EW are comparable to Nafion 115 membrane or even better. However BAM1G membranes showed finite lifetime in an operating fuel cell when run at a constant 500 A/ft 2 at 70˚C in all the longevity evaluations performed with BAM1G, the average time to failure was 350 h. this is particularly low as compared to Nafion ® membranes, which show a lifetime of more than 10,000 h [8,9]. To overcome this problem, a second generation of polymer membrane known as BAM2G was developed.…”