LWN: Comments on "A FUSE implementation for famfs" https://lwn.net/Articles/1020170/ This is a special feed containing comments posted to the individual LWN article titled "A FUSE implementation for famfs". en-us Fri, 03 Oct 2025 01:49:47 +0000 Fri, 03 Oct 2025 01:49:47 +0000 https://www.rssboard.org/rss-specification lwn@lwn.net Non-coherent shared memory? https://lwn.net/Articles/1020963/ https://lwn.net/Articles/1020963/ jagalactic <div class="FormattedComment"> Greg is correct that write order is usually not guaranteed by hardware currently. But when we think about cache coherency it's tempting to always think about the fine-grained multi-writer cases, which are hard - and already have a high performance cost on existing systems. So without sharing disaggregated memory, we already try to avoid these use cases due to the performance cost.<br> <p> But a lot of work flows do what you might call "publish and share", where the publisher needs to flush the processor cache as a last step of publishing, and then there are no hard cache problems because data is not mutated in place while it's shared. This is compatible with a lot of data analytics flows as well as sharing in-memory databases read-only. In fact, it's consistent with how an important subset of databases - the log-structured merge tree (LSM) variants - work, even in writable mode. <br> <p> LSMs write out searchable indexes (often called SST files) which - after the initial write - are read-only until they eventually get merged into new SSTs and get unlinked. Putting SSTs in shared memory (e.g. via famfs) is a pretty nice use case.<br> <p> Mind you, today there are not yet any databases that can directly create files in famfs, due to its pre-allocation requirement - but 1) files can be copied into famfs for sharing, and 2) there are multiple ways famfs and/or apps could adapt to relax famfs' create procedure in future work...<br> <p> </div> Tue, 13 May 2025 14:00:42 +0000 Non-coherent shared memory? https://lwn.net/Articles/1020579/ https://lwn.net/Articles/1020579/ gmprice <div class="FormattedComment"> It's hard because write order isn't guaranteed and explicit flushes that go all the way to the backing media (not just buffers attached to one head) presently don't exist.<br> </div> Fri, 09 May 2025 00:15:57 +0000 Non-coherent shared memory? https://lwn.net/Articles/1020578/ https://lwn.net/Articles/1020578/ DemiMarie <div class="FormattedComment"> Would it make sense to use shared memory in a non-cache-coherent mode with explicit flushes?<br> </div> Thu, 08 May 2025 23:36:00 +0000