[HN Gopher] Transparent memory offloading: more memory at a frac... ___________________________________________________________________ Transparent memory offloading: more memory at a fraction of the cost and power Author : mfiguiere Score : 28 points Date : 2022-06-20 19:44 UTC (3 hours ago) (HTM) web link (engineering.fb.com) (TXT) w3m dump (engineering.fb.com) | throwaway81523 wrote: | Nice, they have reinvented virtual memory and paging to disk. | woleium wrote: | or "Meta re-invents swap disk" | jhgg wrote: | You should read the whole article, because TMO does differ from | traditional swapping. Notably, swapping usually occurs during | periods of extreme memory pressure, where as TMO will offload | memory much sooner and more intelligently. | trhway wrote: | >Using a compressed-memory back end, TMO saves 7 percent to12 | percent of resident memory across five applications. Multiple | applications' data have poor compressibility, such that | offloading to an SSD proves far more effective. | | that in general sounds strange. My experience across different | application types over more than 2 decades has been very | different (not about SSD, i mean the compression in memory have | been always a very efficient approach). | | > Specifically, machine learning models used for Ads prediction | commonly use quantized byte-encoded values that exhibit a | compression ratio of 1.3-1.4x. | | yep, i haven't used compression for that. That though calls for a | question - how many of those models do FB have in memory. Any | chance FB instantiates such a model for every user/request? :) | MarkSweep wrote: | Computer storage hierarchies are sure getting complex with CXL | and NVM in the mix. | | https://en.wikipedia.org/wiki/Compute_Express_Link ___________________________________________________________________ (page generated 2022-06-20 23:00 UTC)