Statistical puzzle

[Garrett Wollman]

Here's something I did not think possible...

Same machine, same disk(s), rewriting a 1 GB file on AFS versus local
disk, excluding close/write-back times.  Inexplicably, writing to
Arla's cache is about 13 Mbyte/s faster than writing directly to the
local file system, and much less jittery.

x /home/wollman/afs.cache.fb
+ /home/wollman/afs.localreal.fb
+--------------------------------------------------------------------------+
|  xxx x                                                        +          |
| xxxx x                                                        +          |
| xxxx x                                                      + +          |
| xxxx xx                                                    ++ ++         |
| xxxxxxx                                                    ++ ++++       |
| xxxxxxx                                                  + +++++++       |
| xxxxxxx                                                 ++ +++++++++     |
|xxxxxxxx                                                 ++++++++++++     |
|xxxxxxxx  x                            +  +++            +++++++++++++    |
|xxxxxxxx xx    ++     +        ++ ++++ +  +++ +++++++++  +++++++++++++++ +|
|  |_A_|                                      |___________A___M_______|    |
+--------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x 100      24.20553     26.263712     24.935205     24.969374    0.40058265
+ 100     27.294477     38.933819     36.572221     35.742469     2.3863808
Difference at 95.0% confidence
        10.7731 +/- 0.474275
        43.1452% +/- 1.89943%
        (Student's t, pooled s = 1.71103)

At a minimum, I think it's safe to say that nnpfs (at least on
FreeBSD) does not put any meaningful impediments in the way of disk
throughput!

-GAWollman