The rationale is an IB RDMA indexed database storing 4 packets per client (thats sufficient for 256 point spectra + indexing / tag overhead , giving a useful profile matching correlation potential.
Each clients meta-data names the sample, the database correlates against each meta matched new inquiry, thus data array alignment is critical to a real-time correlation in a >million stored sample DB. This is all in-memory. No HD access
The IB issue, is how to access raw tables / aligned arrays , can Influx manage this wirespeed correlation more efficiently than building a custom kernel ?
As data structuring is everything, precise targeting is simply keeping a meta-boundary alignment. ie: a dimensional array having client / sample relation… actually having a similar sample alignment no matter who the client…
IB only acts to direct Get http// requests to the node using header directives.
TO what degree does IB inter-operate in keeping alignment in a DB ? None i can see, tho id like to change that.
A basic need to access client Vs sample, the relation between samples (substance) and who the client (owner) of repository. Not that basic with kNN.
This can be held in structure. Who, as in reliability Vs their sample cross-section.
So…getting back to the Infiniband fabric…anyone see a way to address or partition the data intrinsically to allow a per node in-memory search, whilst the IB indexes or positions the arrays immutably in memory (whilst adding each client input, systematically ? We may set each client memory page at 4K to begin, thats only a 4GB in-memory array with a million clients. Our cluster has 1TB. thus its all in directing where each client lays a new egg (4K holds16 samples) before another 4K needs to be assigned. Only scientists exceed 16 samples.
Is this (approx) clear to seek advice how Influx may benefit ?
schema is the drawcard. … Im heading off to the south of France.