ChatGPT reviewed the paper "Detecting Past and Future Change Points in Performance Data"

 

Review of the Paper: Detecting Past and Future Change Points in Performance Data

Pros

1. Comprehensive Introduction to SETDS: The paper provides a detailed explanation of the Statistical Exception and Trend Detection System (SETDS) methodology, which includes statistical filtering, pattern recognition, and IT-control charts. This thorough presentation ensures clarity in understanding the core concepts and their applications.

2. Innovative Use of Exception Values (EVs): The introduction of EVs to measure anomalies quantitatively is a strength. The geometric and statistical interpretation of EVs as indicators of upward or downward trends is well-explained and intuitive.

3. Real-World Application via Perfomalist: The practical implementation of SETDS through the Perfomalist tool demonstrates its usability. Its application to the MongoDB dataset showcases its capability to handle real-world performance data challenges effectively.

4. Extensibility of the Methodology: The paper highlights potential enhancements to the Perfomalist API, such as predicting future change points and integrating critical thresholds, indicating forward-thinking and adaptability.

5. Visualization and Examples: The inclusion of IT-control charts, 2D/3D models, and examples of API inputs and outputs enriches the explanation. It provides a tangible understanding of how the system operates.

6. Focus on Performance Engineering Needs: By addressing anomaly detection, trend forecasting, and capacity management, the paper targets practical and high-priority IT challenges.

---

Cons

1. Limited Comparative Analysis: Although the paper mentions combining SETDS with XGBoost and a voting classifier, it lacks a comprehensive comparison with other existing change point detection methodologies. This limits understanding of its relative advantages and trade-offs.

2. Narrow Scope of Testing: The primary dataset used for evaluation is from MongoDB, which may not represent diverse real-world scenarios. Broader testing across varied datasets and industries could strengthen the claims.

3. Technical Jargon: While thorough, the explanations are heavily technical and may pose a barrier to readers unfamiliar with statistical concepts or performance monitoring.

4. Potential Over-reliance on Assumptions: The use of historical data to establish baselines assumes data stationarity and normality, which may not hold true in all dynamic environments.

5. Limited Exploration of Future Change Point Detection: While the concept of future change point prediction is introduced, its practical implementation and effectiveness remain underexplored.

6. API Accessibility Details: The paper lacks comprehensive information on API performance, scalability, and robustness in high-throughput environments, which are critical for enterprise adoption.

---

Conclusion

The paper presents a robust framework for detecting and forecasting change points in performance data using SETDS and the Perfomalist tool. Its strengths lie in methodological innovation and practical application. However, a deeper evaluation against competing methods, broader testing, and further elaboration on advanced features would enhance its impact and applicability.

"Detecting Past and Future Change Points in Performance Data" - another SETDS paper was accepted ICTDsC 2024 conference in India

The research paper was accepted for ORAL PRESENTATION at ICTDsC 2024 in India.

The abstract is below.

And the paper itself could be found as a preprint at our google drive HERE

My last role model - Prof. Igor Chelpanov

I just came across an article about him, posthumously, in his blessed memory - IN MEMORY OF IGOR BORISOVICH CHELPANOV (По русски) 

It’s interesting that I noticed in this block of mine that he is my last authority HERE, 3 months after his death, which I only found out about now (4 years after..).

He had the greatest talent - Teacher of PhD students. I found him myself after a presentation (about the dynamics of robot grasping devises) made by one of his students, S.N. Kolpashnikov.

This was a turning point in my career (he was my dissertation supervisor) and my entire professional life!

I am immensely grateful to Igor Borisovich and remember him forever!

Meeting of the Department of Automata. Polytech, St. Petersburg, ~1998

In the first row, 1st from left is I. Chelpanov, 2nd from right (I. Trubin)

Besides us: Popov, Krasnoslabodtsev, Dyachenko (head of the department), Volkov (future head of the department) and others

 

 I. B. Chelpanov

"Scale in Clouds. What, How, Where, Why and When to Scale" - my new www.CMG.org presentation

Our presentation (with Jignesh Shah) was accepted for www.CMGimpact.com conference.

Title: Scale in Clouds. What, How, Where, Why and When to Scale

Venue:  Atlanta, GA on February 6 & 7

ABSTRACT:

Scale in Clouds

What, How, Where, Why and When to Scale

Igor Trubin, Jignesh Shah -  Capital One bank 

ABSTRACT

Presentation includes the following discussion themes. 

• What to scale: servers, databases, containers, load balancers.

• How to scale: horizontally/rightsizing, vertically, manually, automatically, ML based, predictive, serverless.

• Where to scale: AWS (ASG,ECS, EKS, ELB), AZURE, GCP, K8s.

• Why to scale: cost optimization, incidents avoidance, seasonality.

• When to scale:  auto-scaling policies and parameters, pre-warming to fight  latency, correlating with business/app drivers.

Presentation includes a user case study of scaling parameters optimization: monitoring, modeling and balancing vertical and horizontal scaling, calculating optimal initial/desired cluster size and more.

AI!?

куда идёт AI

Me at work.... #datamanagement #rap

My 1st dialog with #chatGBT about R (text data to convert to code to create a #dataframe).

ME: I have CSV table  and need a code to convert that to dataframe.

GPT: Sure! Here is some code that will read a CSV file into a dataframe in R:

# set the working directory to the folder where your CSV file is located

setwd("/path/to/folder")

# read the CSV file into a dataframe

mydata <- read.csv("mycsvfile.csv", header = TRUE)

# view the dataframe

mydata

In this code, you will need to replace "/path/to/folder" with the actual path to the folder where your CSV file is located, and "mycsvfile.csv" with the actual name of your CSV file.

The read.csv() function reads a CSV file and creates a dataframe from it. The header = TRUE argument tells R that the first row of the CSV file contains column names.

After running this code, you should have a dataframe called mydata that contains the data from your CSV file. You can view the dataframe by simply typing mydata into the R console.

ME: I know that, but what is I have no ability to use file system? No read.csv, no write.csv, just a coma separated text to code like the following: 

first_column <- c("value_1", "value_2", ...)

second_column <- c("value_1", "value_2", ...)

df <- data.frame(first_column, second_column)

GPT: I see! Here's an example of how you could create a dataframe in R from comma-separated text using the format you provided:

lua

Copy code

# sample comma-separated text

text <- "value_1, value_2, value_3

         value_4, value_5, value_6

         value_7, value_8, value_9"

# split the text into separate lines

lines <- strsplit(text, "\n")[[1]]

# split each line into separate values

values <- lapply(lines, function(x) strsplit(x, ", ")[[1]])

# create a dataframe from the values

df <- data.frame(matrix(unlist(values), nrow=length(lines), byrow=TRUE))

# add column names to the dataframe

colnames(df) <- c("first_column", "second_column", "third_column")

# view the dataframe

In this code, you would need to replace the text variable with your own comma-separated text. The strsplit() function is used to split the text into lines and then into separate values. The lapply() function is used to apply the strsplit() function to each line of the text.

The data.frame() function is used to create a dataframe from the values. The matrix() function is used to reshape the values into a matrix with the appropriate dimensions.

Finally, the colnames() function is used to add column names to the dataframe. You would need to replace the column names in the colnames() function with the appropriate column names for your data.

____

I have tested that and that is not working well.

By googling I got better suggestion: 

https://sparkbyexamples.com/r-programming/different-ways-to-create-a-dataframe-in-r/

SETDS (Statistical Exception and Trend Detection System) vs. AWS/SETDS (Split, Extract, Test, Decide, and Score). Is OpenAI right?

 My SETDS (Statistical Exception and Trend Detection System) method was created about 20 years ago. 

Playing with OpenAI https://chat.openai.com/chat I was giving the following information:

"SETDS (Split, Extract, Test, Decide, and Score) is a method for anomaly detection that was developed by researchers at Amazon Web Services (AWS). The method involves the following steps:

1. Split: The data is split into a training set and a testing set.

2. Extract: Features are extracted from the training set.

3. Test: A model is trained on the training set and tested on the testing set.

4. Decide: Anomaly scores are calculated for each data point in the testing set based on the model's performance.

5. Score: The anomaly scores are used to identify data points that are likely to be anomalous.

The SETDS method is designed to be fast and efficient, and it has been used in a number of applications, including detecting anomalies in log data and in real-time streaming data. It is one of many methods that can be used for anomaly detection, and the suitability of the method will depend on the specific characteristics of the data and the goals of the analysis."

+++++++++++++++++

Interesting that my SETDS also does some kind of scoring by calculating EV - Exception Value. 

Also interesting that I was not able to verify by Google search the information AI bot chat provided....

Cloud Usage Data. Cleansing, Aggregation, Summarization, Interpretability and Usability (#CMGnews) - my presentation

 https://hubs.ly/Q01vGZhc0 

#CMGImpact 2023 conference announcement of the Trubin's presentation about #clouddata

https://youtu.be/WJxdV3jKswg

"#Cloud Usage Data. Cleansing, Aggregation, Summarization, Interpretability and Usability" - CMG Impact'23 presentation (#CMGnews)

My presentation was accepted for CMG Impact'23 (www.CMGimpact.com ) conference (Orlando, FL, Feb. 21-23). 

ABSTRACT:

All cloud objects (EC2, RDS, EBS, ECS/Fargate, K8s, Lambda) are elastic and ephemeral.  It is a real problem to understand, analyze and predict their behavior. But it is really needed for Cost optimization and Capacity management.  The essential requirement to do that is the system performance data. The raw data is collected by observability tools (CloudWatch, DataDog or NewRelic), but it is big and messy.

The presentation is to explain and demonstrate:

- How that should be aggregated and summarize addressing the issue of jumping workload from one cluster to another due to rehydration, releases and failovers.

- How the data should/are to be cleaned by anomaly and change point detection without generating false negatives like seasonality.

- How to summarize the data to avoid sinking in granularity. 

- How to interpret the data to do cost and capacity usage assessments.

- Finally how to use that clean, aggregated and summarized data for Capacity Management by using ML/Predictive analytics.

Hybrid #ChangePointDetection system - #Perfomalist

The paper about using #Perfomalist "Change Point Detection for #MongoDB Time Series Performance Regression" was cited in the following paper: "Estimating Breakpoints in Piecewise Linear Regression Using #MachineLearning Methods", where our method was mentioned as " … offer a hybrid change point detection system..." 

CMG'08 Trip Report

Visualization and Analysis of Performance Data using R Jim Holtman Summary: I did not attend this, but that is about free statistical and graphical tool (“R” tool and “S” language http://www.r-project.org/ ). Note: there is interface to SAS dataset function in open lib: http://lib.stat.cmu.edu/S/dataset Functions that define and manipulate S "dataset" objects. A dataset is a matrix whose columns (variables) may be of different data types. Though motivated by a need to interface to SAS, they are useful in any data analysis. There is some function that relates to SPC: JohnsonSystem (http://lib.stat.cmu.edu/S/JohnsonSystem.q) In 2004 he published CMG paper about R usage: The Use of R for System Performance Analysis . See also Lecture: Graphing in R (http://www.ats.ucla.edu/stat/r/library/lecture_graphing_r.htm) or http://ieee.cincinnati.fuse.net/R_IEEE_V2.pdf Major takeaways: That might be a good SAS/Graph replacement. I also think about writing some "S" program to build SEDS type of Control charts to illustrate how that works, for instance THAT COULD BE USED for a workshop similar Mr. Holtman had done. 

  Automating Process Pathology Detection – Rule Engine Design Hints Ron Kaminski Summary: This is about analytical approach to capture pathologies like run-away and memory leaks. BTW Ron referenced my papers as an example of different (statistical) approach to do the same. This is continuation of his previous work in this field: http://www.cmg.org/proceedings/2003/3027.pdf In private conversation he actually expressed some interest to put together both approaches to see how that works from different angles... I am opened. 

  CMG-T: Modeling and Forecasting Speaker: Dr. Michael A. Salsburg Summary: Just a good an overview and tutorial for queuing theory and simulation based modeling and forecasting vs. statistical modeling-forecasting way I presented in my paper.

  eBay - the Shape of Infrastructure to Come Speaker: Paul Strong Summary: Cloud computing is a “Outsourcing 2.0”, sooner or later even banks will use that approach to use capacity on-demand from cloud instead of having own computer farm…. 

  Exception Based Modeling and Forecasting Speaker: Dr. Igor A. Trubin Summary: This is my presentation which was successful and attracted more than 60 attendees. There were a lot of questions and comments during and before this session, positive comments were received from Mark Friedman (After I had to clarify for him 3-D concept of weekly control charts... - my bad ,I was probably not very clear presenting that...) and Ron Kaminski who expressed some interest in my EV algorithm to capture recent bad trends as that solves some problems of workload pathology recognition on which he has been working recently. 

  So You Want to Manage Your z-Series MIPS? Then Detect & Control Application Workload Variance! Speaker: John S. Van Wagenen, Caterpillar Summary: Unfortunately I could not attend this session as I presented mine in the same time. But this paper is about SEDS-like approach to manage Mainframe capacity! And that presentation got prestigious Mullen award! There is a similar paper written by the same author last year: Performance Monitoring Process for Out of Standard Applications Major takeaways: SEDS approach is valid and our implementation on mainframe might be adjusted using this paper methodology. 

  Predicting the Relative Performance of CPU Speaker: Debbie Sheetz Summary: I used similar approach (see my 1st CMG paper and 1st figure in my last paper) in the past and know how challenging is to apply SPEC or other benchmarks to real servers with different configurations. Major takeaways: This paper could be helpful in соме consolidation projects. 

  Panel: Michelson Panel - Visualization Speaker: Jeff Buzen Summary: That was interesting to see deferent ways to present data visually. During this panel discussions I realized that my weekly control charts and especially 3-D version of that are kind of unique. I have even approached Dr. Buzen, with my comments about that… 

  Mainstream NUMA and the TCP/IP stack Speaker: Mark B. Friedman Summary: This is brilliant but very scary paper. Two scary points: A. For multicore servers the speed of memory access could be unpredictable and sometimes deadly slow because of NUMA – non universal memory access. And there are no any metrics or tools to measure that! B. High performance network (1-10 and higher Gb) cannot be fully utilized, because it might consume all CPU cycles only to process network related interrupts. Major takeaways: It’s OK if network interface bandwidth utilization is low. And we should be careful with using modern multicore processors (8 and more cores). 

  Performance and Capacity Management in an Outsourced Environment Speaker: Jeff Hammond Summary: This is very useful information about what we could expect working with outsourced service (people) or if we got outsourced ourselves. It confirms my own experience. Action Items: Be prepared just in case!

Our poster presentation "SPEC Research — Introducing the #PredictiveAnalytics Working Group" is scheduled at #ICPE2022 #ICPEconf Poster & Demo (Monday - April 11, 2022, 5:15pm)

    https://icpe2022.spec.org/program_files/schedule/

I am happy to co-author 2 papers for #ICPE2022 #ICPEconf

Online conference program  https://icpe2022.spec.org/program_files/schedule/  scheduled our following  presentations:

Poster & Demo (Monday - April 11, 2022, 5:15pm )

André Bauer, Mark Leznik, Md Shahriar Iqbal, Daniel Seybold, Igor Trubin, Benjamin Erb, Jörg Domaschka and Pooyan Jamshidi. SPEC Research — Introducing the Predictive Data Analytics Working Group

Data Challenge (Tuesday - April 12,, 4:15pm - 4:55pm)

Md Shahriar Iqbal, Mark Leznik, Igor Trubin, Arne Lochner, Pooyan Jamshidi and André Bauer. Change Point Detection for MongoDB Time Series Performance Regression

"Change Point Detection (#ChangeDetection) for MongoDB Time Series Performance Regression" paper for ACM/SPEC ICPE 2022 Data Challenge Track

UPDATE: the paper were published - (LINK to PAPER)

The ACM/SPEC ICPE 2022 - Data Challenge Track Committee has decided to ACCEPT our article:

TITLE: Change Point Detection for MongoDB Time Series Performance Regression 
AUTHORS: Md Shahriar Iqbal, Mark Leznik, Igor Trubin, Arne Lochner, Pooyan Jamshidi and André Bauer

https://icpe2022.spec.org/tracks-and-submissions/data-challenge-track/

ABSTRACT

Commits to the MongoDB software repository trigger a collection

of automatically run tests. Here, the identification of commits 

responsible for performance regressions is paramount. Previously, the

process relied on manual inspection of time series graphs to identify

signi￿cant changes, later replaced with a threshold-based detection

system. However, neither system was sufficient for finding changes

in performance in a timely manner. This work describes our recent

implementation of a change point detection system built upon the

Perfomalist approach in combination with XGBoost algorithm. The

algorithm produces a list of change points representing significant

changes from a given history of performance results. We are able

to automatically detect change points and achieve an 83% accuracy,

all while reducing the human effort in the process.

More Perfomalist's  approach details can be found in this blog post:

CPD - Change Point Detection (#ChangeDetection) is implemented in the free web tool Perfomalist

The result of initial usage of Perfomalist CPD API against MongoDB data is published HERE:

Perfomalist #ChangeDetection API was used against #MongoDB #perfomanceTesting dataset

 

 

My Cloud Optimization team at #CapitalOne bank won the CMG.org #Innovation Award (#CMGNews)

  https://www.cmg.org/2022/02/capital-one-announced-as-winner-of-the-impact-innovation-award/

My publications in RG got 5000+ reads

https://www.researchgate.net/profile/Igor-Trubin 

Panel Discussion: Roadmap for Cultivating Performance-Aware Software Engineers

 

"#CloudServers Rightsizing with #Seasonality Adjustments" - my presentation at CMG IMPACT conference (#CMGnews)

Feb 4, 2022 12:15 Virtual at https://cmgimpact.com/sessions-schedule/

"Performance Anomaly and Change Point Detection for Large-Scale System Management" - my paper published at Springer

 

Intelligent Sustainable Systems pp 403-407| Cite as

Performance Anomaly and Change Point Detection for Large-Scale System Management

• Authors
• Authors and affiliations

• Igor Trubin

1. 1.

Conference paper

First Online: 17 December 2021

• 1Downloads

Part of the Lecture Notes in Networks and Systems book series (LNNS, volume 334)

Abstract

The presentation starts with the short overview of the classical statistical process control (SPC)-based anomaly detection techniques and tools including Multivariate Adaptive Statistical Filtering (MASF); Statistical Exception and Trend Detection System (SETDS), Exception Value (EV) meta-metric-based change point detection; control charts; business driven massive prediction and methods of using them to manage large-scale systems such as on-prem servers fleet or massive clouds. Then, the presentation is focused on modern techniques of anomaly and normality detection, such as deep learning and entropy-based anomalous pattern detections.

Keywords

Anomaly detection Change point detection Business driven forecast Control chart Deep Learning Entropy analysis 

This is a preview of subscription content, log in to check access.

References

1. 1.
   Trubin, I.: Exception based modeling and forecasting. In: Proceedings of Computer Measurement Group (2008)Google Scholar
2. 2.
   Jeffrey Buzen, F., Annie Shum, S.: MASF—multivariate adaptive statistical filtering. In: Proceedings of Computer Measurement Group (1995)Google Scholar
3. 3.
   Trubin, I.: Review of IT control chart. CIS J. 4(11), 2079–8407 (2013)Google Scholar
4. 4.
   Perfomalist Homepage, http://www.perfomalist.com. Last accessed on 10 June 2021
5. 5.
   Trubin, I., et al.: Systems and methods for modeling computer resource metrics. US Patent 10,437,697 (2016)Google Scholar
6. 6.
   Trubin, I.: Capturing workload pathology by statistical exception detection. In: Proceedings of Computer Measurement Group (2005)Google Scholar
7. 7.
   Loboz, C.: Quantifying imbalance in computer systems. In: Proceedings of Computer Measurement Group (2011)Google Scholar

Dynamics of Anomalies or Phases in a Dynamic Object Life

A dynamic object may have following several phases in its lifetime:

1.  Initial phase to set a norm - anomalies cannot be detected as there is no baseline sample is established yet. Could be tired later as an outlier.

2.  Stable period without any anomalies.

3. Unstable period when anomalies are appearing: suddenly or with gradually increasing rate.

4. Anomalies are introducing a new norm and the rate of anomalies is gradually decreasing.

5. =>2. The next stable period. 

6. =>3. … and so on.

To detect those dynamic object phases one can use Anomaly and Change Point detection methods. One of them is SETDS (described in this blog), which has been implementing now as a www.Perfomalist.com tool. 

Here is an example how the Perfomalist (Download Input Data Sample) test data is used to detect stable and unstable periods.

Data consists of 28 weeks. To see some dynamic and to  catch when anomalies started appearing, the data was divided into 23 data sets. 

- The 1st one has 4 initial weeks (initial baseline or reference/learning set) plus following week (1st "current" week). 

- The 2nd one has 5 initial weeks as the next (on one week bigger) baseline and following week as the next  "current" week. 

- The 3rd one... the same mechanism as described above.

Then the www.Perfomalist.com was applied 23 times (could be automated using Personalist APIs)  and results were combined into the spreadsheet. 

The table and daily summarized charts are below. The result shows clearly 2nd (stable)  and 3rd (unstable) phases. 

Another way to detect is CPD and to do that another Perfomalist API can be used.

Applying that to the same data the similar result is seen:

The results are similar but a bit different and I know why.... 

Join me with CMG – your technology community – at #CMGIMPACT22. Use code Trubin at cmgimpact.com/ for 50% off IMPACT tickets cmgimpact.com/register/ #cmgnews #technology #InformationTechnology #ITconference #ContinuingEducation #ProfessionalDevelopment

PLEASE REGISTER: https://cmgimpact.com/register/

https://cmgimpact.com/cloud-servers-rightsizing-with-seasonality-adjustments/ 

When the cloud servers rightsizing algorithm calculates the baseline level for the current year application server’s usage, the seasonal adjustment needs to be calculated and applied by adding the additional anticipated change, which could be increasing or decreasing the capacity usage. We describe the method and illustrate it against the real data.

The cloud servers rightsizing recommendation generated based on seasonality adjustments, would reflect the seasonal patterns, and prevent any potential capacity issues or reduce an excess capacity.
The ability to keep multi-year historical data of 4 main subsystems of application servers’ capacity usage opens the opportunity to detect seasonality changes and estimate additional capacity needs for CPU, memory, disk I/Os, and network. A multi-subsystem approach is necessary, as very often the nature of the application could be not CPU but I/Os or Memory or Network-intensive.

Applying the method daily allows downsizing correctly if the peak season passes and the available capacity should be decreased, which is a good way to achieve cost savings.

In the session, the detailed seasonality adjustment method is described and illustrated against the real data. The method is based on and developed by the author’s SETDS methodology, which treats the seasonal variation as an exception (anomaly) and calculates adjustments as variations from a linear trend.

Key Takeaways

• How to build seasonal adjustments into the cloud rightsizing
• To get familiar with cloud objects rightsizing techniques

The Change Point Detection SETDS based method is implemented as a Perfomalist API. Everybody is welcome to test!

How to use it explained HERE:

https://www.trutechdev.com/2021/11/the-change-points-detection-perfomalapi.html

Example of the step jump event detected by the API (Output got from the API call via Postman and spreadsheet was used to chart the result):