| Age | Commit message (Collapse) | Author |
|---|
|
This change removes the distributor and aggregator interfaces in favour
of a single switch interface. Since the switch interface is as powerful
as both the distributor and aggregator, we don't need the latter two.
|
|
|
|
|
|
This change drops the requirement for a clock parameter when
constructing a ComputeMetricExporter, since it will now derive the
timestamp from the recorded metrics.
|
|
This change adds a new API for writing traces in a trace format.
Currently, writing is only supported by the OpenDC VM format, but over
time the other formats will also have support for writing added.
|
|
This change simplifies the TraceFormat SPI interface by reducing the
number of interfaces that implementors need to implement to only
TraceFormat.
|
|
This change updates the ComputeWorkloadLoader to use index column
lookups in order to prevent having to lookup the index for every row.
|
|
This change unifies columns of different tables used by trace formats.
This concretely means that instead of having columns specific per table
(e.g., RESOURCE_ID and RESOURCE_STATE_ID), with this changes these
columns are shared between the tables with a single definition
(RESOURCE_ID).
|
|
This change adds an initial implementation to the trace library for
converting between workload trace formats. Currently the tool supports
only converting to the OpenDC VM trace format. However, in the future,
we will add support for converting between other formats as well.
|
|
This change optimizes the OpenDC VM trace format by removing
unnecessary columns as well as optimizing the writer settings.
The new implementation still supports reading the old trace format in
case users run OpenDC with older workload traces.
|
|
This change adds official support to the trace library for the internal
VM trace format used by OpenDC for its experiments. This is a compact
format that uses Parquet to store the virtual machine trace data in two
Parquet files.
|
|
This change adds support in the trace library for the Azure VM trace
format.
|
|
This change adds support in the trace library for the extended Bitbrains
format. This format is slightly different than the CSV format used by
the original Bitbrains traces and contains more fields.
|
|
This change updates the workload sampling implementation to be more
flexible in the way the workload is constructed. Users can now sample
multiple workloads at the same time using multiple samplers and use them
as a single workload to simulate.
|
|
This change adds support for creating a topology that is CPU-optimized
for simulation. This means that all the CPU resources of a machine are
merged into a single large CPU in order to reduce simulation time.
|
|
This change adds an option for optimizing SimHost simulation by
combining all the CPUs of a machine into a single large CPU. For most
workloads, this does not significantly affect the simulation results,
but does improve the simulation time by a lot.
|
|
This change adds support for creating flexible topologies by creating a
TopologyFactory interface that is responsible for configuring the hosts
of a compute service.
|
|
This change creates a new module for doing simulations with virtual
machine workloads. We have found that a lot of code in the Capelin
experiments code is being re-used by non-experiment modules.
|
|
This change standardizes the metrics emitted by SimHost instances and
their guests based on the OpenTelemetry semantic conventions. We now
also report CPU time as opposed to CPU work as this metric is more
commonly used.
|
|
This change updates the OpenDC compute service implementation with
multiple meters that follow the OpenTelemetry conventions.
|
|
This change refactors the telemetry implementation by creating a
separate MeterProvider per service or host. This means we have to keep
track of multiple metric producers, but that we can attach resource
information to each of the MeterProviders like we would in a real world
scenario.
|
|
This change simplifies the CoroutineMetricReader implementation by
removing the seperation of reader and exporter jobs.
|
|
|
|
|
|
This change moves the fault injection logic directly into the
opendc-compute-simulator module, so that it can operate at a higher
abstraction. In the future, we might again split the module if we can
re-use some of its logic.
|
|
This change updates the compute service to mark servers that cannot be
scheduled as terminated instead of error. Error is instead reserved for
cases where the server is in an error state while running.
|
|
This change fixes an issue in SimHost where guests that where inactive
were also failed, causing an IllegalStateException.
|
|
This change fixes an issue where all servers could not be scheduled due
to the memory size of the host being computed incorrectly.
|
|
This change updates the SimHost implementation to track the up and
downtime of hypervisor guests.
|
|
|
|
This change enables host to overcommit their memory when testing whether
new servers can fit on the host.
|
|
This change adds new metrics for tracking the up and downtime of hosts
due to failures. In addition, this change adds a test to verify whether
the metrics are collected correctly.
|
|
This change adds a metric for the provisioning time of virtual machines
by the compute service.
|
|
This change updates the SimHost implementation to measure the power draw
of the machine without PSU overhead to make the results more realistic.
|
|
This change eliminates unnecessary double to long conversions in the
simulator. Previously, we used longs to denote the amount of work.
However, in the mean time we have switched to doubles in the lower
stack.
|
|
This change upgrades the OpenTelemetry dependency to version 1.5, which
contains various breaking changes in the metrics API.
|
|
This change adds support for failures in the SimHost implementation.
Failing a host will now cause the virtual machine to enter an error
state.
|
|
This change updates the Bitbrains trace tests with the updated trace
that does not hardcode the duration of the trace fragments.
|
|
This change refactors the trace workload in the OpenDC simulator to
track execute a fragment based on the fragment's timestamp. This makes
sure that the trace is replayed identically to the original execution.
|
|
This change updates the FilterScheduler implementation to follow more
closely the scheduler implementation in OpenStack's Nova. We now
normalize the weights, support many of the filters and weights in
OpenStack and support overcommitting resources.
|
|
|
|
This change updates the Kotlin dependencies used by OpenDC to their
latest version.
|
|
This change updates reimplements the performance interference model to
work on top of the universal resource model in
`opendc-simulator-resources`. This enables us to model interference and
performance variability of other resources such as disk or network in
the future.
|
|
This change re-organizes the classes of the compute simulator module to
make a clearer distinction between the hardware, firmware and software
interfaces in this module.
|
|
This change removes the AutoCloseable interface from the
SimResourceProvider and removes the concept of a resource lifecycle.
Instead, resource providers are now either active (running a resource
consumer) or in-active (being idle), which simplifies implementation.
|
|
This change enables the experiments to share the SimResourceInterpreter
across multiple hosts, which allows updates to be scheduled efficiently
for all machines at the same time. This is especially beneficial if the
machines operate on the same time slices.
|
|
This change integrates the power subsystem of the simulator with the
compute subsystem by exposing a new field on a SimBareMetalMachine, psu,
which provides access to the machine's PSU, which in turn can be
connected to a SimPowerOutlet.
|
|
This change splits the functionality present in the CPUFreq subsystem of
the compute simulation. Currently, the DVFS functionality is embedded in
SimBareMetalMachine. However, this functionality should not exist within
the firmware layer of a machine. Instead, the operating system should
perform this logic (in OpenDC this should be the hypervisor).
Furthermore, this change moves the scaling driver into the power
package. The power driver is a machine/firmware specific implementation
that computes the power consumption of a machine.
|
|
This change adds a new interface to the resources library for accessing
metrics of resources such as work, demand and overcommitted work. With
this change, we do not need an implementation specific listener
interface in SimResourceSwitchMaxMin anymore.
Another benefit of this approach is that updates will be scheduled more
efficiently and progress will only be reported once the system has
reached a steady-state for that timestamp.
|
|
This change introduces the SimResourceInterpreter which centralizes the
logic for scheduling and interpreting the communication between resource
consumer and provider.
This approach offers better performance due to avoiding invalidating the
state of the resource context when not necessary. Benchmarks show in the
best case a 5x performance improvement and at worst a 2x improvement.
|
