**Job placement**

- How clever does the job placement need to be?
- Where is the cleverness? i.e. How much in the generic compute layer [site HTCondor] vs. the experiment framework.
- Can we pull some of the scheduling back into the generic compute layer?

Internal compute within the data-lake:

* Assume <50ms fat pipe, does the job placement even need to be topology aware? Just stream on the fast network pipe from anywhere in the lake?


Sites external to the data lake:

* Caches are transparent (or else it won't work) and unmanaged. i.e no explicit data placement done.
* Caches do not expose their contents (?to the job placement systems). Do we ever need to place jobs as a result of their (known, cached) input data?
* If active data set fits in site cache, no clever job placement needed needed.
* If active data set too big for local cache, can we stripe jobs ("A", "B", "C", "D", ..) to ensure that each cache focusses on a fixed set of data (i.e. that does mostly fit in its cache).
* Is it clear from AAA/FAX that a cache miss is OK? i.e. non-fatal.

Workflow-types

* Should separate out our measurements and solutions by workflow type (Raw/Reco, MCgen/Sim/Reco, ESD->AOD, Analysis), since they have different input bandwidth, output bandwidth and cache-ability 


Practical things for the R&D activity:

* Setup suitable HTCondor pool with relevant routes.
	* Configuration depending on needed tags
* Do we do this with a single HTCondor instance? (Even on other sites???). Easy for our site(s) and external cloud.
* How do we want to simulate an "external compute resource"? External site or external cloud?


**Measurement - performance**

* Need to verify the efficiency of physics jobs with different scenarios.  Can we use our standard WLCG job monitoring for this?
	* Efficiency vs network wait for differing caching scenarios, vs. network bandwidth / lossieness
	* Can we benchmark with standardised physics workloads to get a control? HammerCloud.
* How do we measure the cache-hit rate?
* Local node monitoring should just be standard fabric monitoring. What else do we need?