Tofu: A 6D Mash/Torus Interconnect - Fujitsu
Tofu: A 6D Mash/Torus Interconnect - Fujitsu
Tofu: A 6D Mash/Torus Interconnect - Fujitsu
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
1
Highly scalable and usable direct network (<strong>6D</strong> mesh/torus)<br />
10 redundant high BW links, 4 RDMA engines (4x2 simultaneous transfer)<br />
Good collective communication performance with <strong>Tofu</strong> original algorithms<br />
<strong>Tofu</strong> barrier for barrier & reduction in H/W<br />
Direct attached interconnect controller<br />
SPARC64 TM IXfx<br />
CPU<br />
C<br />
Y<br />
Inter- Y<br />
Connect X<br />
B ×<br />
Controller<br />
X<br />
(ICC)<br />
5GB/s x 2<br />
x10 links XYZ<br />
20GB/s x 2 A Z Z B<br />
<strong>Tofu</strong> realizes scalable systems beyond 100,000 nodes<br />
With low power consumption, low latency, and high BW<br />
2<br />
ABC<br />
Copyright 2011 FUJITSU LIMITED
Y+<br />
Z+<br />
X-<br />
Y-<br />
Z-<br />
Y+<br />
Z+<br />
X-<br />
Y-<br />
Z-<br />
Y+<br />
Z+<br />
X-<br />
Y-<br />
Z-<br />
C<br />
C<br />
C<br />
X+<br />
A<br />
X+<br />
A<br />
X+<br />
A<br />
C<br />
C<br />
C<br />
A<br />
A<br />
A<br />
3<br />
B<br />
B<br />
B<br />
X-<br />
Y-<br />
Z+<br />
Z-<br />
B<br />
Y+<br />
C<br />
A<br />
X+
ABC 3D Mesh<br />
X -<br />
Y -<br />
Z+<br />
Z -<br />
4<br />
B<br />
Y+<br />
C<br />
X+<br />
A<br />
XYZ 3D Mesh
xyz<br />
abc<br />
5
abc<br />
6
abc<br />
b→c →a → x → y → z ⇒ a → c → b<br />
abc<br />
7
From To<br />
3 example paths out of 12 possible paths<br />
8
source fault<br />
destination<br />
9
Example)<br />
One application dimension rendered on<br />
two dimensional slice of a machine<br />
0<br />
2<br />
1<br />
4 dimensional slice view<br />
11<br />
3<br />
4<br />
10<br />
10<br />
6<br />
5<br />
2 dimensional slice view<br />
0<br />
2<br />
11<br />
9<br />
7<br />
3<br />
8<br />
10<br />
6<br />
9<br />
1 4 5 8<br />
7
0 11 10 9<br />
0 11 10 9<br />
0 11 1 10 4 9 5 8<br />
0 11 1 10 4 9 5 8<br />
1 4 2 5 3 8 6 7<br />
1 4 2 5 3 8 6<br />
2 3 6 7<br />
7<br />
2<br />
3<br />
6<br />
7<br />
Application X-Axis<br />
The <strong>6D</strong> submesh can be reused.<br />
One of the app-dimensions is<br />
degraded by one hop.<br />
When a node failure occurs,<br />
the running job is force quitted<br />
and restarted from the user’s<br />
checkpoint. 0 10 9<br />
0 10 9<br />
0 10 1 9 4 8 5<br />
0 10 1 9 4 8 5<br />
1 4 2 5 3 6<br />
1 4 2 5 3 6<br />
11<br />
2<br />
2<br />
3<br />
3<br />
6<br />
6<br />
7<br />
7<br />
8<br />
7<br />
Mapping Hole<br />
8<br />
7
Application Y-Axis<br />
2<br />
1<br />
0<br />
2<br />
1<br />
3<br />
4<br />
5<br />
10<br />
3<br />
4<br />
9<br />
6<br />
5<br />
Application X-Axis<br />
the board to be replaced<br />
8<br />
7<br />
mapping hole<br />
mapping hole
13<br />
Copyright 2010 FUJITSU LIMITED