package allocation_manager.proto

Get desktop application:
View/edit binary Protocol Buffers messages

• rpc Allocate (AllocateRequest, AllocateResponse)

  allocation_manager.proto:32

  Allocate attempts to allocate a Unit for this invocation. If successful, invocations will keep the allocation alive by calling Refresh(). If the invocation is instead queued, invocations will continue polling with Allocate() until allocation is successful. Returns: * NOT_FOUND if the Unit is not known to the server * INVALID_ARGUMENT if the request is malformed (see request type for details)

  message AllocateRequest

  allocation_manager.proto:193

  • optional Invocation invocation = 1

    Invocation details for this allocation

  message AllocateResponse

  allocation_manager.proto:201

  • oneof response_type

    • Allocated allocated = 1

      Returned when allocation is successful; the invocation should be able to proceed while simultaneously calling Refresh().

    • Queued queued = 2

      Returned when allocation is unsuccessful due to license contention. The invocation should continue to poll by calling Allocate() passing the invocation_id returned in this message.

• rpc Refresh (RefreshRequest, RefreshResponse)

  allocation_manager.proto:41

  Refresh refreshes a allocation while an invocation is still potentially using the Unit. If the invocation fails to refresh the allocation, the Unit may be allocated to another invocation. Returns: * NOT_FOUND if allocation is not known to the server, and the client should kill the invocation.

  message RefreshRequest

  allocation_manager.proto:254

  Existing invocation to refresh. Must have been allocated by a call to Allocate() successfully (not queued).

  • optional Invocation invocation = 1

    Invocation details for this allocation

  • optional Topology allocated = 3

    Allocated Topology

  message RefreshResponse

  allocation_manager.proto:268

  • string id = 1

    AllocateResponse id that allocated this Unit.

  • optional google.protobuf.Timestamp refresh_deadline = 2

    Time at which the request license will be revoked. The client should issue another RefreshRequest for this invocation_id before this time.

• rpc Release (ReleaseRequest, ReleaseResponse)

  allocation_manager.proto:50

  Release immediately returns an allocated Unit to the pool. This should be called by clients to return Units so they can be quickly allocated to other queued invocations; otherwise, the server will need to wait for the client to time out issuing Refresh calls. Returns: * NOT_FOUND if the allocation is not known to the server.

  message ReleaseRequest

  allocation_manager.proto:277

  • string id = 1

    AllocateResponse that allocated this license.

  message ReleaseResponse

  allocation_manager.proto:282

  Empty response

  (message has no fields)

• rpc Status (StatusRequest, StatusResponse)

  allocation_manager.proto:53

  Status returns all Unit configurations and their status.

  message StatusRequest

  allocation_manager.proto:318

  Empty request

  (message has no fields)

  message StatusResponse

  allocation_manager.proto:322

  • repeated Stats stats = 1

TODO: Fill in ACF fields

Used in: UnitInfo

(message has no fields)

Used in: AllocateResponse

• string id = 1

  Opaque identifier for this invocation, determined by the server. This id should be used by the client in subsequent RefreshRequests for this Unit.

• optional google.protobuf.Timestamp refresh_deadline = 2

  Time at which the request license will be revoked. The client should issue a RefreshRequest for this invocation_id before this time.

• optional Topology topology = 4

  Actual allocated topology

Used in: Status

• ALLOCATION_UNINITIALIZED = 0

• ALLOCATION_ALLOCATED = 1

  Unit currently allocated

• ALLOCATION_PENDING_AVAILABLE = 2

  Unit may be available soon, in deferred release state

• ALLOCATION_AVAILABLE = 3

  Unit not currently allocated

Message used for server config file

OBSOLETED: Replaced with list of TopologyConfigs, below repeated UnitConfig units = 1;

• optional ServerConfig server = 2

• repeated TopologyConfig topology_configs = 3

  List of known Topology configurations

Contains information about a single CPU on a host machine

Used in: HostInfo

• uint32 cpu_idx = 1

  The 0-based idx of the CPU, as seen in /proc/cpuinfo

• uint32 family = 2

  The family code for the CPU (eg. 25 for Zen)

• string model_name = 3

  The model string of the CPU (eg. AMD EPYC 7413 24-Core Processor)

• float freq_mhz = 4

  The speed of the CPU, in MHz (eg. 2650.0)

• uint32 num_cores = 5

  The number of cores had by the CPU (eg. 24)

• string cache = 6

  The amount of cache, represented by a string (eg. 512 KB) TODO: Parse the string properly, so we can have the cache size in int format

Allocates licenses so that they are spread evenly across users. This means that users can jump ahead of the line, or be bumped after new requests.

Used in: TopologyConfig

(message has no fields)

Default prioritizer. Licenses are allocated in the order they are requested.

Used in: TopologyConfig

(message has no fields)

Contains information about a single GPU on a host machine

Used in: HostInfo, UnitInfo

• string bus_id = 1

  The bus_id of the GPU on the host (eg. 46:00.0)

• string gpu_model = 2

  The model string of the GPU (eg. NVIDIA Corporation GP104GL)

• string card_model = 3

  The model string of the GPU's card (eg. Tesla P4)

• uint32 vram_mb = 4

  The amount of video memory available to the GPU, in megabytes

Used in: Status

• HEALTH_UNINITIALIZED = 0

• HEALTH_READY = 1

  Unit is healthy

• HEALTH_UNKNOWN = 2

  Unit health is unknown

• HEALTH_BROKEN = 3

  Unit health is broken/stuck/bad, try automated repair

• HEALTH_SIDELINED = 4

  Unit is being sidelined, do not attempt automated repair

Contains information about a single Host machine

Used in: HostInventory, Topology, UnitInfo

• string hostname = 1

  hostname of the host (i.e. nc-gpu-17.rdu)

• bool reachable = 2

  Indicates whether this host was reachable when the inventory was created

• bool clientd_healthy = 3

  Indicates if the host's bb_clientd service was healthy at inventory creation time

• bool clientd_fuse_ok = 4

  Indicates if the host's bb_clientd_fuse service was healthy at inventory creation time

• repeated GpuInfo gpu_infos = 5

  A list of information about GPUs found on this host (GpuInfo below)

• repeated CpuInfo cpu_infos = 6

  A list of information about CPUs found on this host (CpuInfo below)

Contains the entire inventory of known hosts, along with information gathered about those hosts Map is hostname -> HostInfo (below)

• map<string, HostInfo> hosts = 1

if no fields are provided (empty message), any host will match

Used in: TopologyRequest

• optional string hostname = 1

  if specified, you are requesting a specific host by hostname. Other fields will be ignored

• optional uint32 num_cpus = 2

  if specified, return a host that has the given number of CPUs, or more

• optional uint32 num_gpus = 3

  if specified, return a host that has the given number of GPUs, or more

Used in: AllocateRequest, RefreshRequest

• string id = 1

  Opaque identifier for this invocation, determined by the server. This id should be used by the client in subsequent RefreshRequests for this Unit.

• string owner = 3

  Owning entity issuing the allocation request. Used for logging purposes only. This could be the name of the user or system issuing the request. This must be sent on every Allocate() and Refresh() call in case the server is restarted.

• string purpose = 4

  TODO: for CI use, use the test target so we can estimate run duration. what else can we use this for?

• optional TopologyRequest request = 6

  topology request

Used in: AllocateResponse

• string id = 1

  Opaque identifier for this invocation, determined by the server. This id associates the invocation with a specific spot in the queue; it should be used in subsequent Allocate RPCs or the invocation will be placed at the back of the queue.

• uint32 queue_position = 3

  Position of this invocation in the queue. Invocations in position 1 are next to be allocated, with higher positions getting allocations later than lower positions. The queue_position can increase or decrease over time depending on the prioritization strategy, which may allow users to jump ahead of the line, or be bumped to the end of the line.

• optional google.protobuf.Timestamp next_poll_time = 2

  Time at which client should retry Allocate. The client should issue its next rpc after this time; if it fails to poll for significantly longer (>5s) it may be moved to the back of the queue.

General options for the entire instance

Used in: Config

• uint32 queue_refresh_duration_seconds = 1

  Interval on which actions should refresh their queue position while in queue for an allocation. Default: 15s

• uint32 allocation_refresh_duration_seconds = 2

  Interval on which actions should refresh their allocation while the action is executing. Default: 30s

• uint32 janitor_interval_seconds = 3

  Interval on which to clean up expired/released allocations and queue entries, and promote queued entries to allocations. Default: 1s

• uint32 adoption_duration_seconds = 4

  When the service first starts, for this period of time it holds off on allocating any new Units, and instead listens for and "adopts" allocations that it hears about via the "Refresh" RPC. This duration should be >= allocation_refresh_duration_seconds, to guarantee that all clients Refresh() their allocation before the server moves into the normal operating state. Default: 45s

OBSOLETED: replaced with UnitInfo, below Topology topology = 1;

Used in: StatusResponse

• optional Status status = 2

  Current state of the Unit

• optional google.protobuf.Timestamp timestamp = 4

• optional UnitInfo info = 5

  Unit's info

Used in: Stats

• Health health = 1

• Allocation allocation = 2

• optional google.protobuf.Timestamp mtime = 3

OBSOLETED: replaced with list of HostInfo, etc string config = 1;

Used in: Allocated, RefreshRequest

• string name = 2

  Unique name extracted from the config for speed/convenience.

• repeated HostInfo hosts = 3

  start with just hosts in the topology for first step

Used in: Config

• string name = 1

• repeated string hosts = 2

• oneof prioritizer

  Strategy to distribute Units.

  • FIFOPrioritizer fifo = 3

  • EvenOwnersPrioritizer even_owners = 4

Used in: Invocation

• optional string name = 1

  if specified, you are requesting a known topology by name. Other fields will be ignored

• repeated HostRequest hosts = 2

  if specified, checks for hosts that match the given criteria

A "polymorphic" representation of the various types of info a Unit can hold

Used in: Stats

• oneof Info

  • HostInfo HostInfo = 1

  • GpuInfo GpuInfo = 2

  • AcfInfo AcfInfo = 3