Re: More compatibility for refuse

[Thor Lancelot Simon <tls%panix.com@localhost>]

On Fri, Apr 12, 2019 at 01:55:39AM +0200, Emmanuel Dreyfus wrote:
> 
> On the performance front, it is true that the perfuse stack will cause
> data to be copied back and forth. When I started it, I thought that I
> would quickly have to add shared memory tricks to avoid copying data,
> but it has not been an obvious requirement: both GlusterFS and LTFS have
> decent performances, probably limited by other factors: network latency
> for GlusterFS and LTO drive throughput for LTFS.

I don't know what "decent" is, but, the issue isn't latency; it's
throughput -- for example, https://image.slidesharecdn.com/glusterperformancegdsoct2016-161007213244/95/state-of-gluster-performance-8-638.jpg?cb=1475875989
shows Gluster on Linux hitting 9Gb/sec for sequential I/O.  If we're not in
that neighborhood, I don't think adding additional memory copies is a great
idea.

The other thing to keep in mind is that the system's only got so much memory
bandwidth.  Whatever's consumed by needless copies in the filesystem isn't
available for the rest of the machine's workload.  Admittedly, 9Gb/sec is
only about 10% of the memory bandwidth available on a midrange x86 server
CPU these days; but still, that's not insignificant.

Thor