[HN Gopher] A journey to io_uring, AIO and modern storage devices
       ___________________________________________________________________
        
       A journey to io_uring, AIO and modern storage devices
        
       Author : digikata
       Score  : 70 points
       Date   : 2021-05-24 17:13 UTC (5 hours ago)
        
 (HTM) web link (clickhouse.tech)
 (TXT) w3m dump (clickhouse.tech)
        
       | TristanBall wrote:
       | Everything I've seen say using the no-op scheduler for very low
       | latency devices is the way to go - particularly for nvme and
       | optane devices.. interesting that it doesn't seem to be mentioned
       | here, or in the linked paper?
        
         | wtallis wrote:
         | Recent Linux kernels default to not using an IO scheduler with
         | devices that support multiple queues, which includes NVMe SSDs
         | but not SATA (at least through AHCI HBAs). Individual distros
         | can override the kernel's defaults. However, I'm not sure the
         | difference between no-op and mq-deadline IO scheduling for SATA
         | SSDs would be big enough to matter for the purposes of this
         | article's measurements.
        
           | anarazel wrote:
           | It's a really annoying proxy to determine the default
           | configuration :(. There's plenty devices with multiple
           | channels that benefits from the increased merging possible
           | with a scheduler - particularly around writes, where a lot of
           | consumer SSDs are weak. And conversely, there's plenty
           | workloads on single queue devices that are hurt by the
           | scheduler and some of the other heuristics that trigger for
           | single queue devices...
        
       | infogulch wrote:
       | Pretty interesting to see this and the other io_uring article on
       | the front page today, given yesterday's post,
       | 
       | The Unwritten Contract of Solid State Drives (2017) -
       | https://dl.acm.org/doi/10.1145/3064176.3064187 -
       | https://news.ycombinator.com/item?id=27260522
       | 
       | I'm pretty sure io_uring didn't exist then, and wonder if the new
       | kernel interface would change anything about the paper.
        
       | gumby wrote:
       | I like the xkcd-style graphs. Anybody know what package that is?
        
         | Twirrim wrote:
         | It's a little over-the-top for my taste in this article. Would
         | much rather something a little crisper.
        
         | sz4kerto wrote:
         | https://cran.r-project.org/web/packages/xkcd/vignettes/xkcd-...
        
       ___________________________________________________________________
       (page generated 2021-05-24 23:00 UTC)