Add a resource dimension called "model slots" that bounds the number of models that can share a given model server.

Change: 146167969

Author: Christopher Olston

tensorflow_serving/resources/resource_values.cc

@@ -24,6 +24,7 @@ const char* const kGpu = "gpu";
 }  // namespace device_types
 
 namespace resource_kinds {
+const char* const kNumModelSlots = "num_model_slots";
 const char* const kRamBytes = "ram_in_bytes";
 const char* const kProcessingMillis = "processing_in_millicores";
 }  // namespace resource_kinds

tensorflow_serving/resources/resource_values.h

@@ -26,7 +26,8 @@ namespace serving {
 // Standard device types.
 namespace device_types {
 
-// CPU(s) and main memory.
+// The primary devices such as CPU(s) and main memory, as well as aspects of the
+// server as a whole.
 extern const char* const kMain;
 
 // Graphics processing unit(s).
@@ -37,6 +38,10 @@ extern const char* const kGpu;
 // Standard resource kinds.
 namespace resource_kinds {
 
+// If a server can accommodate at most N models, depicted as the server having N
+// "model slots", this is the number of slots needed or allocated.
+extern const char* const kNumModelSlots;
+
 // RAM in bytes.
 extern const char* const kRamBytes;