Difference between revisions of "Enhanced SOLE for AI"

From IHE Wiki
Jump to navigation Jump to search
Line 25: Line 25:
  
 
A study is acquired. An Orchestrator/Task Manager invokes one or more applications.
 
A study is acquired. An Orchestrator/Task Manager invokes one or more applications.
I need to know:
+
I need to:
 
*Right-size the resources for an AI Application
 
*Right-size the resources for an AI Application
*Allocate the resource allocation to optimize the algorithm performance
+
*Allocate the resources to optimize the algorithm performance
*How much CPU, what kind of accelerator, memory, bandwidth storage  
+
*Decide how much CPU, accelerator capacity, memory, bandwidth, and storage are needed
*What is the uptime, latency, requests/algorithm utilization
+
*Determine the current uptime, response latency, request volume, and algorithm utilization  
*How to troubleshoot a problem
+
*Troubleshoot a problem
*Licensing deployment model assessment
+
*Assess model licensing
  
  

Revision as of 10:29, 1 July 2022

Enhanced SOLE for AI Power Point presentation

1. Proposed Workitem: Enhanced SOLE for AI

  • Proposal Editor: Chris Lindop, Neil Tenenholtz, Rob Horn, Brian Bialecki
  • Editor: <Name of candidate Lead Editor for the Profile, if known>
  • Date: N/A (Wiki keeps history)
  • Version: N/A (Wiki keeps history)
  • Domain: <Domain name Radiology

2. The Problem

Healthcare providers have a strong desire to increase throughput and efficiency, both to improve the quality and timeliness of care and to control costs which include:

  • Clinical workflow restructuring and optimization (existing in SOLE)
  • Disaster support and readiness (existing in SOLE)
  • Resource utilization and bottleneck identification (extended)
  • Resource allocation for algorithms and workflows (new)
  • Fault identification and troubleshooting (new)


The SOLE profile currently lacks the capability to address the complexity introduced by AI and distributed, cloud-native applications. This profile extends SOLE to address these gaps.

3. Key Use Case

A study is acquired. An Orchestrator/Task Manager invokes one or more applications. I need to:

  • Right-size the resources for an AI Application
  • Allocate the resources to optimize the algorithm performance
  • Decide how much CPU, accelerator capacity, memory, bandwidth, and storage are needed
  • Determine the current uptime, response latency, request volume, and algorithm utilization
  • Troubleshoot a problem
  • Assess model licensing


'Should I pay for a machine running 24 hours or on-demand running?'

4. Standards and Systems

SOLE, Standardize Operational Logging of Events

/ IHE_RAD_Suppl_SOLE.pdf

W3C Trace Context

/ www.w3.org/TR/trace-context-1 (V1, status: recommendation)
/ w3c.github.io/trace-context (v2, status: draft)

5. Discussion

  1. This proposal is a result of an ad hoc committee formed within IHE/ACR to address “CP-458 SOLE: Add events to support AIR and AIW-I”. #Detailed proposed changes were initially discussed as part of the IHE Maintenance and to be included in this workitem.
  2. With many proprietary methods available today and W3C developing the common standard, the timing is good for IHE to converge on promoting a single industry-wide method that benefits community at large.
  3. Technical Approach is to include the existing actors/transactions in SOLE with enhancements that address the AI workflow and used case outlined earlier.

Risks, Open Issues

  1. What to do about non-http transports. (out-of-scope?)