One of several elements that separates high performance computing GPUs from their gaming and graphics brethren is the addition of ECC codes, which target critical bit-flip errors in memory, which can lead to invalid results or system problems.
While ECC is often deemed a necessary component for confirming the viability of simulation results, it does come with a performance price. According to a team of researchers from the San Diego Supercomputer Center and Los Alamos National Laboratory, enabling ECC cuts the size of the system available by 10% because of the amount of memory consumed by the error correction codes. They note that additionally, turning ECC on “reduces simulation speed, resulting in greater opportunity for other sources of error such as disk failures in large file systems, power glitches, and unexplained node failures during the timeframe of the calculation.”
With this performance and greater potential for failure in mind, the question turns to whether ECC is preventing enough critical flaws to justify its price. In other words, are these errors so common that ECC is necessary? As one might imagine, this is a difficult question to tackle since the compute time with multiple hardware, application, GPU and other issues are all involved. However, the team took the question of ECC usefulness across large XSEDE systems, including Keeneland at the Georgia Institute of Technology, a smaller production cluster at Los Alamos, and on Dante at SDSC, which is equipped with GPUs of the gaming variety (so without any ECC).
As seen in the graph, the performance penalty on Keeneland, which was the largest system used in the test GPU/node count-wise, is certainly observable. Similar results in terms of performance hits were observed on the other systems as well. But most interestingly, when it came to actually seeing how useful the ECC was overall for all of the systems, it turned out that there were very few errors—and in fact, the most significant errors or problems with the results when compared across the different systems were based on the hardware itself, faulty motherboards or other variables…not the types the errors ECC is designed to address—at least for the AMBER molecular dynamics code that was used as the basis for the cross-system testing. There is far more detail about the nature of this MD code and why it was particularly relevant for this sort of testing in the full paper.
As the researchers summarize, “Although the ability of ECC to detect and correct single bit errors is undeniably useful in theory, the practical application of this technology may not be in the interests of the MD community.” They point to the rarity of ECC correctable errors and note that they do “not outweigh the costs in terms of system size and calculation speed,” noting that “the errors appear to be so rare in production GPU calculations that their rate of incidence could not be quantified in this experiment.”
Finally, they surmise that overall, “the fact that other sources of hardware error were observed during the experiment, regardless of ECC status, indicates that there are much more probable ways for simulations to fail and that such failures most likely cause the simulation to crash rather than to produce bad data.”
Again, there is a great deal more information in the full piece, but this sparks new life in the debate over whether or not ECC is all it’s cracked up to be for some scientific applications. Does this mean a new life for low-brow gaming graphics cards in large-scale scientific computing sites? Probably not—but an interesting read nonetheless.