Chalmers University of Technology and Hebrew University of Jerusalem is seeking collaboration on the topic of cache memory hierarchies. Both groups have been doing research, individually, on investigating novel locality models so as to improve cache performance. We have had informal, email-based discussions on this topic over the last few months and now want to join forces.

We are seeking support to meet at the cluster meetings arranged regularly every three months for the next 12 months.

The particular topics we will jointly work on is described below:

Caching is based on the principle of locality, often expressed by
means of a computation's working set. But some of the data may not exhibit significant locality, thereby degrading the performance of the caches. This motivates the design of a bypass mechanism, that will identify such data and prevent it from entering the cache.

Analysis of SPEC benchmarks shows that typically 85-95% of the memroy references are targetted at only 5-15% of the addresses. This motivates the definition of a "core" working set, i.e. those addresses that are most highly used. At Hebrew University we are developing techniques to identify the core working set at runtime, based on access characteristics, with the goal of being able to cache core items while bypassing non-core items. The main topics of the joint research are the mechanisms for identifying core items, and the cache structures that exploit this distinction.

Other ideas that have been pursued at Chalmers is to explore alternative cache fetch policies that only selectively install blocks in the cache if they are expected to be reused. Recently we proposed a technique by which bypassing is applied to the last-level cache by using the upper-level caches as a buffer for bypassed blocks. This technique turned out to be quite effective.

Chalmers and Hebrew now want to pursue this topic together by exchanging ideas based on respective experiences in the field.

We are seeking funding for traveling to work on the topics above. We will meet regularly (at the clusters meetings). We are concretely seeking funding for traveling to these meetings (four trips over the next 12 months and two persons each).

Estimated cost per trip: 2 x 1000 = 2000 EURO
Cost for four trips: 8000 EURO