Sampling API

The new API replaces the flexible but irreducibly inefficient scheme where the next sample time for a sampling was determined by the return value of the sampler callback.

Definitions

probe
A location or component of a cell that is available for monitoring.
sample
A record of data corresponding to the value at a specific probe at a specific time.
sampler
A function or function object that receives a sequence of sample records.
schedule
A function or function object that, given a time interval, returns a list of sample times within that interval.

Probes

Probes are specified in the recipe objects that are used to initialize a simulation; the specification of the item or value that is subjected to a probe will be specific to a particular cell type.

using probe_tag = int;

struct probe_info {
    probe_tag tag;         // opaque key, returned in sample record
    any address;           // cell-type specific location info

    template <typename X>
    probe_info(X&& a, probe_tag tag = 0):
       tag(tag), address(std::forward<X>(x)) {}
};

std::vector<probe_info> recipe::get_probes(cell_gid_type gid);

The tag field has no semantics for the engine. It is provided merely as a way of passing additional metadata about a probe to any sampler that polls it, with a view to samplers that handle multiple probes, possibly with different value types.

Probe addresses are decoupled from the cell descriptions themselves — this allows a recipe implementation to construct probes independently of the cells themselves. It is the responsibility of a cell group implementation to parse the probe address objects wrapped in the any address field.

The _k_th element of the vector returned by get_probes(gid) is identified with a probe-id: cell_member_type{gid, k}.

One probe address may describe more than one concrete probe, depending upon the interpretation of the probe address by the cell group. In this instance, each of the concerete probes will be associated with the same probe-id. Samplers can distinguish between different probes with the same id by their probe index (see below).

Samplers and sample records

Data collected from probes (according to a schedule described below) will be passed to a sampler function or function object:

struct probe_metadata {
    cell_member_type id; // probe id
    probe_tag tag;       // probe tag associated with id
    unsigned index;      // index of probe source within those supplied by probe id
    util::any_ptr meta;  // probe-specific metadata
};

using sampler_function =
    std::function<void (probe_metadta, size_t, const sample_record*)>;

where the parameters are respectively the probe metadata, the number of samples, and finally a pointer to the sequence of sample records.

The probe_id, identifies the probe by its probe-id (see above).

The probe_tag in the metadata is the key given in the probe_info returned by the recipe.

The index identifies which of the possibly multiple probes associated with the probe-id is the source of the samples.

The any_ptr value iin the metadata points to const probe-specific metadata; the type of the metadata will depend upon the probe address specified in the probe_info provided by the recipe.

One sample_record struct contains one sample of the probe data at a given simulation time point:

struct sample_record {
    time_type time;    // simulation time of sample
    any_ptr data;      // sample data
};

The data field points to the sample data, wrapped in any_ptr for type-checked access. The exact representation will depend on the nature of the object that is being probed, but it should depend only on the cell type and probe address.

The data pointed to by data, and the sample records themselves, are only guaranteed to be valid for the duration of the call to the sampler function. A simple sampler implementation for double data, assuming one probe per probe id, might be as follows:

using sample_data = std::map<cell_member_type, std::vector<std::pair<double, double>>>;

struct scalar_sampler {
    sample_data& samples;

    explicit scalar_sample(sample_data& samples): samples(samples) {}

    void operator()(probe_metadata pm, size_t n, const sample_record* records) {
        for (size_t i=0; i<n; ++i) {
            const auto& rec = records[i];

            const double* data = any_cast<const double*>(rec.data);
            assert(data);
            samples[pm.id].emplace_back(rec.time, *data);
        }
    }
};

The use of any_ptr allows type-checked access to the sample data, which may differ in type from probe to probe.

Model and cell group interface

Polling rates, policies and sampler functions are set through the simulation interface, after construction from a recipe.

using sampler_association_handle = std::size_t;
using cell_member_predicate = std::function<bool (cell_member_type)>;

sampler_association_handle simulation::add_sampler(
    cell_member_predicate probe_ids,
    schedule sched,
    sampler_function fn,
    sampling_policy policy = sampling_policy::lax);

void simulation::remove_sampler(sampler_association_handle);

void simulation::remove_all_samplers();

Multiple samplers can then be associated with the same probe locations. The handle returned is only used for managing the lifetime of the association. The cell_member_predicate parameter defines the set of probe ids in terms of a membership test.

Two helper functions are provided for making cell_member_predicate objects:

// Match all probe ids.
cell_member_predicate all_probes = [](cell_member_type pid) { return true; };

// Match just one probe id.
cell_member_predicate one_probe(cell_member_type pid) {
    return [pid](cell_member_type x) { return pid==x; };
}

The sampling_policy policy is used to modify sampling behaviour: by default, the lax policy is to perform a best-effort sampling that minimizes sampling overhead and which will not change the numerical behaviour of the simulation. The exact policy requests that samples are provided for the exact time specified in the schedule, even if this means disrupting the course of the simulation. Other policies may be implemented in the future, but cell groups are in general not required to support any policy other than lax.

The simulation object will pass on the sampler setting request to the cell group that owns the given probe id. The cell_group interface will be correspondingly extended:

void cell_group::add_sampler(sampler_association_handle h, cell_member_predicate probe_ids, sample_schedule sched, sampler_function fn, sampling_policy policy);

void cell_group::remove_sampler(sampler_association_handle);

void cell_group::remove_all_samplers();

Cell groups will invoke the corresponding sampler function directly, and may aggregate multiple samples with the same probe id in one call to the sampler. Calls to the sampler are synchronous, in the sense that processing of the cell group state does not proceed while the sampler function is being executed, but the times of the samples given to the sampler will typically precede the time corresponding to the current state of the cell group. It should be expected that this difference in time should be no greater the the duration of the integration period (i.e. mindelay/2).

If a cell group does not support a given sampling_policy, it should raise an exception. All cell groups should support the lax policy, if they support probes at all.

Schedules

Schedules represent a non-negative, monotonically increasing sequence of time points, and are used to specify the sampling schedule in any given association of a sampler function to a set of probes.

A schedule object has two methods:

void schedule::reset();

time_event_span events(time_type t0, time_type t1)

A time_event_span is a std::pair of pointers const time_type*, representing a view into an internally maintained collection of generated time values.

The events(t0, t1) method returns a view of monotonically increasing time values in the half-open interval [t0, t1). Successive calls to events — without an intervening call to reset() — must request strictly subsequent intervals.

The data represented by the returned time_event_span view is valid for the lifetime of the schedule object, and is invalidated by any subsequent call to reset() or events().

The reset() method resets the state such that events can be retrieved from again from time zero. A schedule that is reset must then produce the same sequence of time points, that is, it must exhibit repeatable and deterministic behaviour.

The schedule object itself uses type-erasure to wrap any schedule implementation class, which can be any copy–constructable class that provides the methods reset() and events(t0, t1) above. Three schedule implementations are provided by the engine:

// Schedule at integer multiples of dt:
schedule regular_schedule(time_type dt);

// Schedule at a predetermined (sorted) sequence of times:
template <typename Seq>
schedule explicit_schedule(const Seq& seq);

// Schedule according to Poisson process with lambda = 1/mean_dt
template <typename RandomNumberEngine>
schedule poisson_schedule(time_type mean_dt, const RandomNumberEngine& rng);

The schedule class and its implementations are found in schedule.hpp.

Helper classes for probe/sampler management

The simulation and mc_cell_group classes use classes defined in scheduler_map.hpp to simplify the management of sampler–probe associations and probe metdata.

sampler_association_map wraps an unordered_map between sampler association handles and tuples (schedule, sampler, probe set, policy), with thread-safe accessors.

Batched sampling in mc_cell_group

The fvm_multicell implementations for CPU and GPU simulation of multi-compartment cable neurons perform sampling in a batched manner: when their integration is initialized, they take a sequence of sample_event objects which are used to populate an implementation-specific multi_event_stream that describes for each cell the sample times and what to sample over the integration interval.

When an integration step for a cell covers a sample event on that cell, the sample is satisfied with the value from the cell state at the beginning of the time step, after any postsynaptic spike events have been delivered.

It is the responsibility of the mc_cell_group::advance() method to create the sample events from the entries of its sampler_association_map, and to dispatch the sampled values to the sampler callbacks after the integration is complete. Given an association tuple (schedule, sampler, probe set, policy) where the schedule has (non-zero) n sample times in the current integration interval, the mc_cell_group will call the sampler callback once for probe in probe set, with n sample values.

In addition to the lax sampling policy, mc_cell_group supports the exact policy. Integration steps will be shortened such that any sample times associated with an exact policy can be satisfied precisely.