[tahoe-dev] warning: Boing-Boinging imminent

[Greg Troxel]

"Zooko O'Whielacronx" <zooko at zooko.com> writes:

> Last time I checked there were two servers on the Test Grid that had
> latencies > 1 second for every request. They were the ones named
> "sunpal" and "linuxpal". Yep, I see that they are still connected:
> http://pubgrid.tahoe-lafs.org/

Those are mine.   I have no good ideas about why they are slow.
Does the server status page give any relevant performance stats?

I just restarted linuxpal (a historical name, it's running netbsd-5, and
tahoe 1.7.1).

two things which are non-default but should be normal:

  Those machines do have lease expiration turned on.  Perhaps that causes
  slowness?  (I believe that servers without expiration just don't make
  sense, but maybe I'm confused.)

  They also have a reserved value set to avoid using all space.

Is there a way to do 'tahoe debug fetch-share serverid storageid' and
isolate the operations?  That kind of fine-grained ability would help a
lot with performance testing.


> This partly shows how the Test Grid is serving two incompatible purposes:
>
> Purpose 1: demonstrating the behavior and performance of Tahoe-LAFS to
> curious newcomers
>
> Purpose 2: experimenting with how Tahoe-LAFS handles an uncontrolled
> and dynamic set of servers, and failure and degraded conditions
>
> Now that Tahoe-LAFS v1.8.0 is out the door I intend to experiment with
> setting up a new "Demo Grid" which is high performance and high
> reliability. Demo Grid could serve purpose 1 and Test Grid could serve
> purpose 2.

It might be that splitting them reduces the avlue of purpose 2 and the
result is net worse than where we are now.  An alternative approach is
to view things that make the pubgrid unusable as high-priority things to
fix.

> Atlas Networks and Rackspace have each contributed some servers to our
> open source project, and I'm thinking of maybe setting up the Demo
> Grid to be comprised solely of servers operated by those two
> companies.

Or you could put a few more boxes into pubgrid server status.  My
biggest pubgrid problem is running out of servers to upload shares, at
least with 3/7/10.  Having 12 non-full responding servers would be a
huge increase.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 194 bytes
Desc: not available
URL: <http://tahoe-lafs.org/pipermail/tahoe-dev/attachments/20100929/07b2c7fe/attachment.pgp>