{"id":2610,"date":"2021-10-18T09:00:27","date_gmt":"2021-10-18T00:00:27","guid":{"rendered":"https:\/\/jennifersoft.com\/en\/uncategorized\/%ec%8a%a4%ed%83%9d%ed%8a%b8%eb%a0%88%ec%9d%b4%ec%8a%a4-%ec%83%98%ed%94%8c%eb%a7%81%ec%9d%84-%ec%9d%b4%ec%9a%a9%ed%95%9c-%ec%84%b1%eb%8a%a5-%eb%b6%84%ec%84%9d\/"},"modified":"2022-04-18T15:51:20","modified_gmt":"2022-04-18T06:51:20","slug":"2021-10-23","status":"publish","type":"post","link":"https:\/\/jennifersoft.com\/en\/blog\/tech\/2021-10-23\/","title":{"rendered":"Stacktrace Sampling Based Performance Analysis"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nThe J company that\ndeveloper Y works for is using an old version of installation type of Jira\nsoftware\n\n\n\n<\/p>\n\n\n\n

Although the developer knows that it is installed somewhere in IDC and the equipment is very old, it is very difficult to either upgrade it or install it in a new location due to the license policy that requires allocation of license on a hardware basis.<\/p>\n\n\n\n


Developer Y with anxiety about the response time that goes down very much sometimes (Why the hell is this slow!)
Developer Y tries to install an APM that the company is developing to do a method profiling, but that attempt fails miserably very soon.<\/p>\n\n\n\n

\u201cWhich\npackage should I configure?\u201d<\/p>\n\n\n\n

\u201cLet\u2019s\nmonitor it first!\u201d<\/p>\n\n\n\n

Using\nthe X-View chart, which is the key function of APM installed, developer Y can\ncheck the transaction located at the top of the Y axis (Oh, yes..).<\/p>\n\n\n\n

\"\"
Transaction marked at the top of Y axis in the X-View chart<\/figcaption><\/figure>\n\n\n\n

In\nhigh spirits, developer Y tries to drag the point that is about to penetrate\nthe ceiling of the chart.<\/p>\n\n\n\n

\"\"
 There is some Not Profiled area that lasts for more than 12 seconds<\/figcaption><\/figure>\n\n\n\n

\u201cIt\ntakes more than 12 seconds to finish the transaction… What is the reason?\u201d<\/p>\n\n\n\n

The area of delay is not a traditional bottleneck such as an SQL or HTTP call, so developer Y feels a heavy heart.
Knowing that Jira is developed by the company called Atlassian,
so developer Y decides to configure the profile of the com.atlassian package.<\/p>\n\n\n\n

so developer Y decides to configure the profile of the com.atlassian package.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Urghh.. As soon as it is configured, the system CPU usage rate exceeds 52%.
After calming down the worrisome heart, developer Y drags the X-View chart.<\/p>\n\n\n\n

Can\ndeveloper Y find out why it is happening?<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Ah..\nAlthough numerous methods are profiled, unfortunately, most of methods take 0ms\nto 1ms. Looking at the X-View chart, the service response time slows down a\nlot.<\/p>\n\n\n\n

\"\"<\/figure><\/div>\n\n\n\n

The manager might blame developer Y. (Come on! Cheer up!)<\/p>\n\n\n\n

In\nthis case, is it a bad attempt to do profiling?
\nIf a method of certain class within the configured package caused a delay, it\nwould have been possible to find out the cause.<\/p>\n\n\n\n

However,\nin a modern application, very bad performance for a specific method developed\nby users does not impact the response time very much.
\nBecause in most environments, it is developed by using a well prepared\nframework.
\nBy profiling the com.atlassian package, it is not possible to check the\nsignificant profile information.
\nIn most cases, caches are prepared internally, and it is not very highly likely\nto cause a performance problem.<\/p>\n\n\n\n

If\nthere is an SQL or HTTP call that takes long time, then you might want to tune\nthe target at the remote site, but if there is not, then it would be much more\ndifficult to resolve.
\nWhat if you configure all the packages corresponding to the loaded class(!!)\nMaybe we can find it.
\nBut, due to the amount of additional load and data applied to the application,\nwe will have to pay enormously high costs.<\/p>\n\n\n\n

Alright, then how\nshould we analyze the situation?<\/p>\n\n\n\n

There\nare two ways of performance analysis – profiling and sampling.<\/p>\n\n\n\n

1. Profiling
<\/h4>\n\n\n\n

It\ninvolves collecting performance data from each method that runs and finding and\nimproving the methods that take a long time.<\/p>\n\n\n\n

\"\"
Track the beginning and end of each method and measure the response time.
The slowest methods for Transaction1 and Transaction2 are Method2 and Method 4, respectively.
<\/figcaption><\/figure>\n\n\n\n

2. Sampleing<\/h4>\n\n\n\n

It\ninvolves collecting performance data from each method that runs and finding and\nimproving the methods that take a long time.<\/p>\n\n\n\n

\"\"
Periodically collect snapshots of methods running in each thread and measure the response time.
In the figure, we can tell that Method1 is the cause of delay.
<\/figcaption><\/figure>\n\n\n\n

Profiling\nis a traditional method, so it has been used for a long time throughout\nhistory.
\nThus, lots of monitoring products basically provide profiling oriented\nfunctions.
\nIt sounds near perfect, but actually, it has some limitations as follows.<\/p>\n\n\n\n

  1. A user must be aware of the target\n method to monitor (Maybe developers would know).<\/li>
  2. If used for applications in the\n operation environment with high traffic, the performance can be affected\n (What if I configure the package with many classes, oops?).<\/li><\/ol>\n\n\n\n

    However,\ntraffic leading to true performance improvement occurs in the operation\nenvironment only.
    \nIf there is a way to monitor them in any kind of environment, then it would be\npossible to overcome the limitation of profiling?
    \nOne of the methods is called sampling<\/strong>\nthat we discussed earlier.
    \nThis method involves finding the location of bottleneck based on the frequency\nof appearance by periodically collecting stacktraces from each thread where the\nmethod call is generated, without directly tracking individual methods.
    \n(When the stacktrace is collected 3 times for the time period of 100ms, if\nmethod a appears all 3 times, then we assume that method a is running for\n300ms.)
    \nCompared to the profiling based method, it is less accurate, but it relatively\nless affects the performance while collecting data.<\/p>\n\n\n\n

    What\nare the main causes of processing delays?<\/p>\n\n\n\n