[Background] ============= Erofs already supports chunk deduplication across different images to minimize disk usage since v6.1. Furthermore, we can make inodes among different images share page cache for these deduplicated chunks to reduce the memory usage. This shall be much usable in container scenarios as deduplication is requisite for container images. [Implementation] ================ This is achieved by managing page cache of deduplicated chunks in blob's address space. In this way, all inodes sharing the deduplicated chunk will refer to and share the page cache in the blob's address space. [Restriction] ============== The page cache sharing feature also supports .mmap(). The reverse mapping requires that one vma can not be shared among inodes and can be linked to only one inode. As the vma will be finally linked to the blob's address space when page cache sharing enabled, the restriction of the reverse mapping actually requires that the mapped file area can not be mapped to multiple blobs. Thus page cache sharing can only be enabled for those files mapped to one blob. The chunk based data layout guarantees that a chunk will not cross the device (blob) boundary. Thus in chunk based data layout, those files smaller than the chunk size shall be guaranteed to be mapped to one blob. As chunk size is tunable at a per-file basis, this restriction can be relaxed at image building phase. As long as we ensure that the file can not be deduplicated, the file's chunk size can be set to a reasonable value larger than the file size, so that the file contains only one chunk, in which case page cache sharing feature can be enabled on this file later. [Effect] ======== The final optimization result of this feature depends on the following factors: 1. The number of deduplicated (shared) chunks. Images sharing most of the layers (e.g. a base image and v1 image based on the base image) will achieve better optimization. 2. As the restriction mentioned above, the number of files for which page cache sharing can ben enabled among the files accessed. I test the workload of starting up Tensorflow, which will access quite many (~5K) files among the startup phase. I get the base image of Tensorflow from [1] and build a new image (e.g. v1 image) on top of this base image. Since the image got from [1] is in OCI format, I have to convert it to erofs format with buildkit[2], with default chunk size of 1MB. I run containers from these two images with containerd (base image first, v2 image secondly). The (page cache) memory usage of the rootfs (container image) is shown as below: | page cache sharing | page cache sharing | disabled | enabled ------------------------|-----------------------|------------------- First container | | page cache usage (MB) | 150 | 150 ------------------------+-----------------------|------------------- Second container | | page cache usage (MB) | 150 | 7 It can be seen that most (~95%, 143MB/150MB) memory usage reduced under this workload (when starting following containers sharing container image layers). The remained 7MB memory usage is consumed by directories, since page cache sharing is enabled only for regular files in this RFC implementation. [1] docker.io/tensorflow/tensorflow:2.10.0 [2] https://github.com/moby/buildkit Jingbo Xu (6): erofs: remove unused device mapping in the meta routine erofs: unify anonymous inodes for blob erofs: alloc anonymous file for blob in share domain mode erofs: implement .read_iter for page cache sharing erofs: implement .mmap for page cache sharing erofs: enable page cache sharing in fscache mode fs/erofs/fscache.c | 231 ++++++++++++++++++++++++++++++++------------ fs/erofs/inode.c | 27 ++++++ fs/erofs/internal.h | 8 +- fs/erofs/super.c | 2 + 4 files changed, 202 insertions(+), 66 deletions(-) -- 2.19.1.6.gb485710b
