Prefetcher

Prefetcher

The objective of the Prefetcher is to promote and demote content which is expected to be used in the near future or frequently. The prefetcher only applies to data which is already staged within Hermes. In order to activate prefetching, a Prefetcher Trait can be attached to a Tag (or Bucket) to indicate that prefetching should be enabled and which kind of prefetching should be applied.

Usage (Configuration)

To enable prefetching with configuration.

Usage (C++)

To enable prefetching, attach the Prefetcher trait to a tag (e.g., Bucket). In this example, we attach the PrefetcherTrait with the kApriori prefetcher constant to the SimulationBucket, which represents the data for the simulation workload.

auto bkt = HERMES->GetBucket("SimulationBucket");
TagId bkt_id = bkt.GetId();
hermes::TraitId apriori_trait =
    HERMES->RegisterTrait<hermes::PrefetcherTrait>(
      "apriori", hermes::PrefetcherType::kApriori);
TraitId trait_id = HERMES->GetTraitId("apriori");
bkt.AttachTrait(apriori_trait);

Application Tracing

In order to support prefetching, we implement a tracing system within Hermes. The tracer is called for every Put and Get operation within Hermes. It stores the information called for the Put or Get internally within a multiple-producer single-consumer (MPSC) shared-memory queue, which is asynchronously digested by the prefetcher.

The tracer collects the following information:

1. Operation (Put or Get)
2. Blob Id
3. Bucket Id
4. Blob Size
5. Timestamp (from program start)
6. Rank (if MPI)

In the binary file, we store the following information:

1. Operation (Put or Get)
2. Blob Name (64-bit Hash)
3. Bucket Name (64-bit Hash)
4. Blob Size (64-bit)
5. Timestamp (from program start)
6. Rank (if MPI)

The trace can be analyzed live within the prefetcher or be stored in a binary log (created using Cereal) and analyzed offline.

Apriori Prefetcher

Currently, we implement the apriori prefetcher. Many applications exhibit completely deterministic I/O patterns. Deep Learning applications for example will have the same I/O pattern when the randomness seed is fixed and all other paramters remain the same. Many HPC workloads are executed repeatedly with the same parameters for reasons such as reproducability. This prefetcher assumes that the user will supply a schema with the exact momements of when to promote and demote content.

Future Work

1. Live Prefetcher: Use some sort of short-term memory models to analyze I/O statistics generated live.