John Groves, Micron Micron recently released the first RFC for famfs [1]. Although famfs is not CXL-specific in any way, it aims to enable hosts to share data sets in shared memory (such as CXL) by providing a memory-mappable fs-dax file system interface to the memory. Sharable disaggregated memory already exists in the lab, and will be possible in the wild soon. Famfs aims to do the following: * Provide an access method that provides isolation between files, and does not tempt developers to mmap all the memory writable on every host. * Provide an an access method that can be used by unmodified apps. Without something like famfs, enabling the use of sharable memory will involve the temptation to do things that may destabilize systems, such as mapping large shared, writable global memory ranges and hooking allocators to use it (potentially sacrificing isolation), and forcing the same virtual address ranges in every host/process (compromising security). The most obvious candidate app categories are data analytics and data lakes. Both make heavy use of "zero-copy" data frames - column oriented data that is laid out for efficient use via (MAP_SHARED) mmap. Moreover, these use case categories are generally driven by python code that wrangles data into appropriate data frames - making it straightforward to put the data frames into famfs. Furthermore, these use cases usually involve the shared data being read-only during computation or query jobs - meaning they are often free of cache coherency concerns. Workloads such as these often deal with data sets that are too large to fit in a single server's memory, so the data gets sharded - requiring movement via a network. Sharded apps also sometimes have to do expensive reshuffling - moving data to nodes with available compute resources. Avoiding the sharding overheads by accessing such data sets in disaggregated shared memory looks promising to make make better use of memory and compute resources, and by effectively de-duplicating data sets in memory. About sharable memory * Shared memory is pmem-like, in that hosts will connect in order to access pre-existing contents * Onlining sharable memory as system-ram is nonsense; system-ram gets zeroed... * CXL 3 provides for optionally-supported hardware-managed cache coherency * But "multiple-readers, no writers" use cases don't need hardware support for coherency * CXL 3.1 dynamic capacity devices (DCDs) should be thought of as devices with an allocator built in. * When sharable capacity is allocated, each host that has access will see a /dev/dax device that can be found by the "tag" of the allocation. The tag is just a uuid. * CXL 3.1 also allows the capacity associated with any allocated tag to be provided to each host (or host group) as either writable or read-only. About famfs Famfs is an append-only log-structured file system that places many limits on what can be done. This allows famfs to tolerate clients with a stale copy of metadata. All memory allocation and log maintenance is performed from user space, but file extent lists are cached in the kernel for fast fault resolution. The current limitations are fairly extreme, but many can be relaxed by writing more code, managing Byzantine generals, etc. ;) A famfs-enabled kernel can be cloned at [3], and the user space repo can be cloned at [4]. Even with major functional limitations in its current form (e.g. famfs does not currently support deleting files), it is sufficient to use in data analytics workloads - in which you 1) create a famfs file system, 2) dump data sets into it, 3) run clustered jobs that consume the shared data sets, and 4) dismount and deallocate the memory containing the file system. Famfs Open Issues * Volatile CXL memory is exposed as character dax devices; the famfs patch set adds the iomap API, which is required for fs-dax but until now missing from character dax. * (/dev/pmem devices are block, and support the iomap api for fs-dax file systems) * /dev/pmem devices can be converted to /dev/dax mode, but native /dev/dax devices cannot be converted to pmem mode. * /dev/dax devices lack the iomap api that fs-dax uses with pmem, so the famfs patch set adds that. * VFS layer hooks for a file system on a character device may be needed. * Famfs has uncovered some previously latent bugs in the /dev/dax mmap machinery that probably require attention. * Famfs currently works with either pmem or devdax devices, but our inclination is to drop pmem support to, reduce the complexity of supporting two different underlying device types - particularly since famfs is not intended for actual pmem. Required :- Dan Williams Christian Brauner Jonathan Cameron Dave Hansen [LSF/MM + BPF ATTEND] I am the author of the famfs file system. Famfs was first introduced at LPC 2023 [2]. I'm also Micron's voting member on the Software and Systems Working Group (SSWG) of the CXL Consortium, and a co-author of the CXL 3.1 specification. References [1] https://lore.kernel.org/linux-fsdevel/cover.1708709155.git.john@xxxxxxxxxx/#t [2] https://lpc.events/event/17/contributions/1455/ [3] https://www.computeexpresslink.org/download-the-specification [4] https://github.com/cxl-micron-reskit/famfs-linux Best regards, John Groves Micron
