Add pseudo-LRU eviction policy to the Configuration Cache #9
Comments
seldridge
added a commit
that referenced
this issue
May 6, 2016
This corrects two problems that were preventing multi-transaction
workloads from working:
* The learning variant of DANA's Transaction Table was using the
`numOutputs` to determine when all data had been written to the
output queue of the X-FILES Arbiter for feedforward transactions.
For feedforward transactions `numOutputs` is never set. Feedforward
transactions now use `nodesInCurrentLayer` to determine when all
data has been written and that transaction is finished.
* The ASID--NNID Table Walker was never transitioning back to an idle
state after servicing a request by the Configuration Cache to load a
specific neural network configuration.
This enables multiple transactions using different neural network
configurations to be run back to back. This has been tested with serial
transactions that do not require eviction of neural network
configurations from the Configuration Cache.
Multiple transactions that require eviction from the Configuration Cache
will, however, fail (#9).
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
The original C++ simulator implemented a pseudo-LRU replacement algorithm for replacing unused entries (specifically a clock LRU). Add this back in as the current way that this is working is a "first unused" which will kill the possibility of temporal locality.
The text was updated successfully, but these errors were encountered: