Browse Prior Art Database

Innovative Instrumentation Method for Monitoring Latency Bands Between Transactions and Obtaining Outlier Trace

IP.com Disclosure Number: IPCOM000246169D
Publication Date: 2016-May-12
Document File: 2 page(s) / 44K

Publishing Venue

The IP.com Prior Art Database

Abstract

Described is a method for monitoring latency bands between transactions and obtaining outlier trace.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 52% of the total text.

Page 01 of 2

Innovative Instrumentation Method for Monitoring Latency Bands Between Transactions and Obtaining Outlier Trace

In today's data center, every service request response must have a consistent and low latency, or the overall operation response time will be tragically slow. The overall response time is dominated by the long tail distribution of the parallel operations. While today, there are latency monitoring events in the Performance Monitoring Counters (PMU) that accumulate the latency counts and thus gives an average latency. Average latency numbers may be misleading many times as few dangerously high (or unacceptable ) latency numbers will be hidden in averaging with other readings. Since it is not possible to get those extreme cases, and also not possible to get to know how many such extreme cases (outliers) have occurred, one needs a different approach than what is being done today (averaging). Also in case of high latency, there is no way today to assess who could be contributing to the higher latency.

    A new proposal to monitor latencies that fall under predefined latency periods (bands), which is under user control, setup events that tracks predefined latency bands and count number of such occurrences and divert each of them into their respective buckets for counting. By bucketing, it is meant setting up of four latency bands that one wishes to monitor or track. Ex: Whenever a latency event counts number cycles, it took between "req" and "resp", the design compares latency numbers/cycle counts with the predefined bands or water mark and allocates that to a particular bucket as defined by the user. Also, by monitoring the highest latency level counts and dumping the traces of selected events into an Events_Trace register,one gets possible contributors to the higher latencies (see the figure below).

Latency threshold monitor proposal:A new proposal to monitor latencies that fall under predefined latency periods (bands)

Setup events that tracks predefined l...