miswiring models for autonomic naming in a data center fabric
TRANSCRIPT
Miswiring models for autonomic naming in a data center fabric
Keith Marzullo, UC San Diego Dept of CSEwith Pardis Miri, Meg Walraed-Sullivan, and Amin Vahdat
... and others of the PortLand project
Sunday, January 16, 2011
Architecture
with three levels and k port switches: 5k2/4 switches, k3/4 hostsSunday, January 16, 2011
UNRP: Autonomic naming
L3
L2
L1
Sunday, January 16, 2011
UNRP: Autonomic naming
L2
Sunday, January 16, 2011
UNRP: Autonomic naming
22 55 11 44
21 31 32 23
Sunday, January 16, 2011
Miswiring
Sunday, January 16, 2011
Modeling miswirings
First idea: use a threshold modelAssume graphs are d-deficient:
In n × n bipartite graph, each node has at least n − d edges
Sunday, January 16, 2011
The good and bad with this model
k-deficiency is a nice mathematical abstractionn and k are large, and so a large d should indicate many failures => unlikelydiameter of bipartite graphs are easily bounded in terms of d-deficiency
... but we rejected ithard to instantiateissues of fabric startup
Sunday, January 16, 2011
Model we used
Assume l levels; hosts connected to one (lowest) level
Assign addresses quickly and using few messagesThe closer to a fat tree, the better the addressesTwo hosts separated by more than 2 l −1 won’t be able to communicate with each otherDo diagnostics separately
Transfer risk of loss of reliability from instantiation to quality of immediate result
Reminiscent of graceful degradation
Sunday, January 16, 2011
In action
Sunday, January 16, 2011
In action
L2 L2L2 L2L2 L2L2
L3 L3 L2 L2
L1L1 L1L1 L1L1 L1L1
L1
Sunday, January 16, 2011
In action
L2 L2L2 L2L2 L2L2
L3 L3 L2 L2
L1L1 L1L1 L1L1 L1L1
L1
Sunday, January 16, 2011
In action
L2 L2L2 L2 L2L2
L3 L3
L1L1 L1L1 L1 L1L1L1
Sunday, January 16, 2011
In action
2 51 4 66
21 21,1 1 21
Sunday, January 16, 2011
In action
2 51 4 66
21 21,1 1 21
Sunday, January 16, 2011
Current work and questions
Allowing for hosts to be connected at any level
Level assignment becomes harderMiswiring model still appropriate?
DiagnosticsCan we be more precise about the closer to a fat tree, the better the addresses?
Sunday, January 16, 2011