Re: [Lsf-pc] [LSF/MM ATTEND] [LSF/MM TOPIC] NOVA: A log-structured file system for NVM

[Andiry Xu <jix024@eng.ucsd.edu>]

On Wed, Feb 10, 2016 at 3:14 AM, Jan Kara <jack@suse.cz> wrote:
>   Hi,
>
> On Tue 09-02-16 22:15:00, Andiry Xu wrote:
>> I would like to attend LSF/MM to present a new file system design for
>> non-volatile main memories.
>
> Well, LSF/MM is based more on discussions than presentations. Also a lot of
> people from LSF/MM will go to Vault so I don't think presenting at LSF/MM
> would make much sense. But we register your attend request.
>
>                                                                 Honza
>

Thank you Jan.


>> The goal of NOVA is to provide high performance and strong consistency
>> (atomic metadata, data, mmap updates) at the same time.
>>
>> Although NOVA is a LFS, as it targets a different memory technology,
>> it looks very different from conventional LFS. Some of the key design
>> decisions of NOVA include:
>>
>> Per-inode log: Each inode has a log, with tail pointer points to the
>> latest committed entry. This allows for high concurrency and parallel
>> log scanning.
>>
>> Log is a linked list: Allows for fine-grained, page-size granularity
>> garbage collection. Also, allocating log space is easy since NOVA does
>> not need to allocate large, contiguous space for logs.
>>
>> Cheap atomicity: To atomically append a log, NOVA simply appends the
>> log entry and then updates the log tail pointer. To coordinate updates
>> across logs, NOVA appends each log and journal the log tails to update
>> all logs atomically. This mechanism has lower overhead than journaling
>> (which doubles writes) and shadow paging (which causes cascade
>> updates).
>>
>> Stores file data outside the log: NOVA appends metadata of writes to
>> the log and handles file data in a copy-on-write way. This results in
>> a shorter log, and garbage collection is simple and efficient, as NOVA
>> never copies data in log cleaning.
>>
>> Highly scalable: NOVA has free list, journal and inode table at each
>> CPU to avoid global locking and scalability bottlenecks.
>>
>> NOVA is available on GitHub:
>>
>> https://github.com/NVSL/NOVA
>>
>> And the paper of NOVA will appear in FAST 2016. I will also give a
>> talk in the Vault conference.
>>
>> I'm also interested in any NVM/DAX related topics. Thank you.
> --
> Jan Kara <jack@suse.com>
> SUSE Labs, CR