[tahoe-dev] Benchmarks - 1.7.1 vs. 1.8.0c3

Kyle Markley kyle at arbyte.us
Tue Sep 7 00:11:27 UTC 2010


>> Is there a way to have tahoe pick a web.port instead of me specifying
>> it? Or a better idea about what I could do?
> Yeah, if you set web.port to "0", it will pick an unused port, and write
> it into node.url, so CLI commands will work as usual. It won't read the
> port out of node.url (or persist it in any other way), so it'll change
> from one run to the next, but I'd bet that's sufficient for the
> benchmarking tests.

Excellent!  That makes it easy for me to start up lots of storage nodes.

I've finally gotten around to running some benchmarks with several remote
storage nodes.  This is a 4-machine configuration:

(A) is a WinXP-32 laptop on wired LAN
(B) is a Vista-32 laptop on wired LAN
(C) is a Win7-64 laptop on wireless
(D) is an OpenBSD desktop on wired LAN

(A) and (B) were mostly idle.  (C) was lightly loaded (web browsing and
e-mail).  (D) is my buildbot and was running the tahoe coverage tests
during the benchmark.  As if that wasn't enough interference, a deep-check
operation or three were also running on my real grid (not the benchmarking
grid), affecting all these machines.

Each machine ran 10 temporary storage nodes and the client used the
default 3/7/10 encoding.  The client ran on (D) and ran my typical sequence
of uploading large files (4 * 16MiB) and small files (1024 * 64KiB) then
downloading the same.

Because each machine ran 10 storage nodes, I'm completely relying on
randomness to distribute the shares across the several machines. 
Variability between runs may be high.

I ran the whole experiment twice, the second time after the buildbot and
deep-check activity had subsided.

Results are wall-clock times in seconds and are legible in a monospace

1.7.1:              1      2
 upload large     79.51  69.38
 upload small    906.16 797.78
 download large   86.53 115.46
 download small  262.99 273.89

1.8.0c3:            1      2
 upload large     54.13  42.52
 upload small    731.62 727.82
 download large  148.01 211.20
 download small  205.67 215.79

1.7.1 is consistently faster downloading large files.
1.8.0c3 is consistently faster downloading small files.
These characterizations may only be valid for a low-latency grid.

Kyle Markley
-------------- next part --------------
A non-text attachment was scrubbed...
Name: tahoe-bench.pl
Type: application/octet-stream
Size: 13473 bytes
Desc: not available
URL: <http://tahoe-lafs.org/pipermail/tahoe-dev/attachments/20100906/ecd1bbe9/attachment-0001.obj>

More information about the tahoe-dev mailing list