JVM tuning ideas

JVM Operation forecast

    Mainly depends on jstat gc -pid Command Statistics GC Operation of the . For example, the growth rate of young generation objects 、Young GC The trigger frequency and time consumption of each time 、 Every time Young GC After that, how many people survive and enter the elderly generation 、Full GC The trigger frequency and time consumption of each time etc. . On the basis of this information jvm Parameter adjustment .

The rate at which young people grow

Can execute orders jstat -gc pid 1000 10 ( every other 1 Seconds to perform 1 subcommand , In total 10 Time ), Through observation EU(eden The use of the zone ) To estimate... Per second eden About how many new objects , If the system load is not high , You can put the frequency 1 Second to second 1 minute , even to the extent that 10 Minutes to see the whole situation . Be careful , General systems may have peak and daily periods , So we need to estimate the growth rate of objects in different situations at different times .

Young GC The trigger frequency and time consumption of each time

If we know the growth rate of young people, we can infer that eden The size of the area is calculated as Young GC How often does it trigger ,Young GC The average time spent by YGCT/YGC The formula works out , Based on the results, we can probably know how long the system will last because Young GC How long is it going to be stuck .

Every time Young GC After that, how many people survive and enter the elderly generation

This is because I knew about it before Young GC The frequency of , Suppose every 5 Minutes at a time , Then you can execute the command jstat -gc pid 300000 10 , Observe the results every time eden,survivor And changes in the use of older people , In every time gc after eden The use of the zone is generally greatly reduced ,survivor And the elderly are likely to grow , The object of these increases is every time Young GC After the survival of the object , And you can see that every time Young GC How many objects are there in the old age , So we can calculate the growth rate of objects in the old age .

Full GC The trigger frequency and time consumption of each time

Knowing the growth rate of objects in the old age, we can calculate Full GC The trigger frequency of ,Full GC You can use the formula for each time FGCT/FGC calculated .

The theoretical basis of optimization

     In fact, it's simply ：

•     If an object can be eliminated in the younger generation, try not to let it enter the older generation
•     Try not to trigger full GC

  Trigger young GC Conditions ：

•    eden Lack of space

  Trigger full GC Conditions ：

•   There is not enough space in the old age
•   There is not enough space for permanent generation
•   young GC when Triggered the old age space allocation guarantee mechanism .

  Conditions for the object to enter the old age ( That is, it will affect the growth rate of objects in the old age )：

•   Large objects go directly into the old generation
•   Older people enter the older generation
•   young GC According to the dynamic age determination mechanism , Put... At one time Survivor Object is greater than Survivor District 50%, Then the extra objects will be put into the old generation

full gc Than minor gc More reasons

• Excess caused by insufficient meta space full gc
• Display call System.gc() Cause superfluous full gc, This kind of general online as far as possible ­XX:+DisableExplicitGC Parameter disable , If you add this JVM Launch parameters , So the code is called System.gc() No effect
• The guarantee mechanism of space distribution in the old days

Analysis methods

1. First , According to the object growth rate in the old age , Come to the conclusion that in theory full GC The frequency of . Compare the theoretical frequency with the actual frequency .
2. If the actual frequency is much higher than the theoretical frequency , So according to full gc Than minor gc There are many reasons to analyze what kind of reasons . For example, try to increase the size of meta space 、 Ban System.gc() And avoid triggering the space allocation guarantee mechanism of the elderly generation .
3. If the example frequency is not much different from the theoretical frequency , It shows that the problem lies in the object growth rate in the old age . At this point, there are two ways to analyze ：

         3.1   Consider reducing the rate at which objects enter the old age . Use thread analysis to determine whether a large number of large objects are generated , Whether in the younger generation young GC It triggers the dynamic age determination mechanism to make the object enter the old age .

         3.2  Consider increasing the heap size , To reduce full GC frequency .