Remèdes aux oomkill, warm-ups, et lenteurs pour des conteneurs JVM 

Orateurs Brice Dutheil @BriceDutheil Jean-Philippe Bempel @jpbempel 

Agenda My container gets oomkilled How does the memory actually work Some case in hands Container gets respawned Things that slow down startup Break 

The containers are restarting. What’s going on ? $ kubectl get pods NAME READY STATUS RESTARTS AGE my-pod-5759f56c55-cjv57 3/3 Running 7 3d1h 

The containers are restarting. What’s going on ? On Kubernetes, one should inspect the suspicious pod $ kubectl describe pod my-pod-5759f56c55-cjv57 ... State: Running Started: Mon, 06 Jun 2020 13:39:40 +0200 Last State: Terminated Reason: OOMKilled Exit Code: 137 Started: Thu, 06 Jun 2020 09:20:21 +0200 Finished: Mon, 06 Jun 2020 13:39:38 +0200 

My container gets oomkilled 

🔥 🔥 🔥🔥 🔥 🔥 🔥 🚨 Crisis mode 🚨 If containers are oomkilled Just increase the container memory limits and investigate later 

Monitor and setup alerting 

Monitor the oomkills In Kubernetes cluster monitor the terminations with metrics ● kube_pod_container_status_last_terminated_reason, if the exit code is 137, the attached reason label will be set to OOMKilled ● Trigger an alert by coupling with kube_pod_container_status_restarts_total 

Monitor the resident memory of a process RSS Heap Max Heap Liveset memory limit 

Monitor the resident memory of a process Depending on the telemetry libraries (eg Micrometer) you may have those ● Heap Max : jvm_memory_max_bytes ● Heap Live : jvm_memory_bytes_used ● Process RSS : process_memory_rss_bytes And system ones, eg Kubernetes metrics ● Container RSS : container_memory_rss ● Memory limit : kube_pod_container_resource_limits_memory_bytes 

💡 Pay attention to the unit Difference between 1 MB and 1 MiB ? 

💡 Pay attention to the unit The SI notation, decimal based : 1 MB reads as megabyte and means 1000² bytes The IEC notation, binary based : 1 MiB reads as mebibyte and means 1024² bytes ⚠ The JVM uses the binary notation, but uses the legacy units KB, MB, etc. OS command line tools generally use the binary notation. https://en.wikipedia.org/wiki/Binary_prefix#/media/File:Binaryvdecimal.svg At gigabyte scale the difference is almost 7% 1GB ≃ 0.93 GiB 

Oomkilled ? 

Oomkilled ? Is it a memory leak ? Is it misconfiguration ? 

Linux Oomkiller ● Out Of Memory Killer Linux mechanism employed to kill processes when the memory is critically low ● For regular processes the oomkiller selects the bad ones. ● Within a restrained container, i.e. with memory limits, ○ If available memory reaches 0 in this container then the oomkiller terminates all processes ○ There is usually a single process in container 

Linux oomkiller Oomkills can be reproduced synthetically docker run --memory-swap=100m --memory=100m \ --rm -it azul/zulu-openjdk:11 \ java -Xms100m -XX:+AlwaysPreTouch --version 

Linux oomkiller And in the system logs $ tail -50 -f $HOME/Library/Containers/com.docker.docker/Data/log/vm/console.log ... [ 6744.445271] java invoked oom-killer: gfp_mask=0xcc0(GFP_KERNEL), order=0, oom_score_adj=0 ... [ 6744.451951] Memory cgroup out of memory: Killed process 4379 (java) total-vm:3106656kB, anon-rss:100844kB, file-rss:15252kB, shmem-rss:0kB, UID:0 pgtables:432kB oom_score_adj:0 [ 6744.473995] oom_reaper: reaped process 4379 (java), now anon-rss:0kB, file-rss:32kB, shmem-rss:0kB ... 

Oomkilled ? Is it a memory leak ? or … Is it misconfiguration ? 

Memory of a process 

Memory of a JVM process 

Memory As JVM based developers ● Used to think about JVM Heap sizing, mostly -Xms, -Xmx, … ● Possibly some deployment use container-aware flags: -XX:MinRAMPercentage, -XX:MaxRAMPercentage, … JVM Heap Xmx or MaxRAMPercentage 

Why should I be concerned by native? ● But ignoring other memory zones 💡 Referred to as native memory, Or on the JVM as off-heap memory 

Why should I be concerned by native? JDK9 landed container support! Still need to need to make the cross multiplication yourself. 

Why should I be concerned by native? Still have no idea what is happening off-heap JVM Heap ❓ 

Why should I be concerned by native? Still have no idea what is happening off-heap JVM Heap https://giphy.com/gifs/bitcoin-crypto-blockchain-trN9ht5RlE3Dcwavg2 

If you don’t know what’s there, … How can you size properly the heap or the container ? JVM Heap 

JVM Memory Breakdown Running A JVM requires memory: ● The Java Heap 

JVM Memory Breakdown Running A JVM requires memory: ● The Java Heap ● … 

JVM Memory Breakdown Running A JVM requires memory: ● The Java Heap ● The Meta Space (pre-JDK 8 the Permanent Generation) ● … 

JVM Memory Breakdown Running A JVM requires memory: ● The Java Heap ● The Meta Space (pre-JDK 8 the Permanent Generation) ● Direct byte buffers ● Code cache (compiled code) ● Garbage Collector (like card table) ● Compiler (C1/C2) ● Symbols ● etc. 

JVM Memory Breakdown Running A JVM requires memory: ● The Java Heap ● The Meta Space (pre-JDK 8 the Permanent Generation) ● Direct byte buffers ● Code cache (compiled code) ● Garbage Collector (like card table) ● Compiler (C1/C2) ● Threads ● Symbols ● etc. JVM subsystems 

JVM Memory Breakdown Except a few flags for meta space, code cache, or direct memory There’s no control over memory consumption of the other components But It is possible to get their size at runtime. 

Let’s try first to monitor 

Let’s try first to monitor Eg with micrometer time series ● jvm_memory_used_bytes ● jvm_memory_committed_bytes ● jvm_memory_max_bytes Dimensions ● area : heap or nonheap ● id : memory zone, depends on GC and JVM jvm_memory_used_bytes{area="nonheap",id="CodeHeap 'profiled nmethods'",} 8231168.0 jvm_memory_used_bytes{area="heap",id="G1 Survivor Space",} 5242880.0 jvm_memory_used_bytes{area="heap",id="G1 Old Gen",} 1.164288E7 jvm_memory_used_bytes{area="nonheap",id="Metaspace",} 4.180964E7 jvm_memory_used_bytes{area="nonheap",id="CodeHeap 'non-nmethods'",} 1233536.0 jvm_memory_used_bytes{area="heap",id="G1 Eden Space",} 1.2582912E7 jvm_memory_used_bytes{area="nonheap",id="Compressed Class Space",} 5207416.0 jvm_memory_used_bytes{area="nonheap",id="CodeHeap 'non-profiled nmethods'",} 1590528.0 

Let’s try first to monitor Don’t forget the JVM native buffers ● jvm_buffer_total_capacity_bytes 

Monitor them 

Monitor them RSS k8s memory limit Heap max 

Monitor them Eden Old gen 

Monitor them JVM off-heap pools Not really practical to look at 

Monitor them 💡Stack the pools (if supported by your observability tool) Missing Data RSS Stacked pools 

Monitoring is only as good as data is there Observability metrics rely on MBean to get memory areas Most JVM don’t export metrics for everything that uses memory 

Time to investigate the footprint With diagnostic tools 

RSS is the real footprint $ ps o pid,rss -p $(pidof java) PID RSS 6 4701120 

jcmd – a swiss knife Who knows ? Who used it already ? 

jcmd – a swiss knife Get the actual flag values $ jcmd $(pidof java) VM.flags | tr ' ' '\n' 6: ... -XX:InitialHeapSize=4563402752 -XX:InitialRAMPercentage=85.000000 -XX:MarkStackSize=4194304 -XX:MaxHeapSize=4563402752 -XX:MaxNewSize=2736783360 -XX:MaxRAMPercentage=85.000000 -XX:MinHeapDeltaBytes=2097152 -XX:NativeMemoryTracking=summary ... PID Xms Xmx 

JVM’s Native Memory Tracking 1. Start the JVM with -XX:NativeMemoryTracking=summary 2. Later run jcmd $(pidof java) VM.native_memory Modes ● summary ● detail ● baseline/ diff 

$ jcmd $(pidof java) VM.native_memory 6: Native Memory Tracking: Total: reserved=7168324KB, committed=5380868KB - Java Heap (reserved=4456448KB, committed=4456448KB) (mmap: reserved=4456448KB, committed=4456448KB) - Class (reserved=1195628KB, committed=165788KB) (classes #28431) ( instance classes #26792, array classes #1639) (malloc=5740KB #87822) (mmap: reserved=1189888KB, committed=160048KB) ( Metadata: ) ( reserved=141312KB, committed=139876KB) ( used=135945KB) ( free=3931KB) ( waste=0KB =0.00%) ( Class space:) ( reserved=1048576KB, committed=20172KB) ( used=17864KB) ( free=2308KB) ( waste=0KB =0.00%) - Thread (reserved=696395KB, committed=85455KB) (thread #674) (stack: reserved=692812KB, committed=81872KB) (malloc=2432KB #4046) (arena=1150KB #1347) - Code (reserved=251877KB, committed=105201KB) (malloc=4189KB #11718) (mmap: reserved=247688KB, committed=101012KB) - GC (reserved=230739KB, committed=230739KB) (malloc=32031KB #63631) (mmap: reserved=198708KB, committed=198708KB) - Compiler (reserved=5914KB, committed=5914KB) (malloc=6143KB #3281) (arena=180KB #5) - Internal (reserved=24460KB, committed=24460KB) (malloc=24460KB #13140) - Other (reserved=267034KB, committed=267034KB) 

$ jcmd $(pidof java) VM.native_memory 6: Native Memory Tracking: Total: reserved=7168324KB, committed=5380868KB - Java Heap (reserved=4456448KB, committed=4456448KB) (mmap: reserved=4456448KB, committed=4456448KB) - Class (reserved=1195628KB, committed=165788KB) (classes #28431) ( instance classes #26792, array classes #1639) (malloc=5740KB #87822) (mmap: reserved=1189888KB, committed=160048KB) ( Metadata: ) ( reserved=141312KB, committed=139876KB) ( used=135945KB) ( free=3931KB) ( waste=0KB =0.00%) ( Class space:) ( reserved=1048576KB, committed=20172KB) ( used=17864KB) ( free=2308KB) ( waste=0KB =0.00%) - Thread (reserved=696395KB, committed=85455KB) (thread #674) (stack: reserved=692812KB, committed=81872KB) (malloc=2432KB #4046) (arena=1150KB #1347) - Code (reserved=251877KB, committed=105201KB) (malloc=4189KB #11718) (mmap: reserved=247688KB, committed=101012KB) - GC (reserved=230739KB, committed=230739KB) (malloc=32031KB #63631) (mmap: reserved=198708KB, committed=198708KB) - Compiler (reserved=5914KB, committed=5914KB) (malloc=6143KB #3281) (arena=180KB #5) - Internal (reserved=24460KB, committed=24460KB) (malloc=24460KB #13140) - Other (reserved=267034KB, committed=267034KB) (classes #28431) (thread #674) Java Heap (reserved=4456448KB, committed=4456448KB) 

$ jcmd $(pidof java) VM.native_memory 6: Native Memory Tracking: Total: reserved=7168324KB, committed=5380868KB - Java Heap (reserved=4456448KB, committed=4456448KB) (mmap: reserved=4456448KB, committed=4456448KB) - Class (reserved=1195628KB, committed=165788KB) (classes #28431) ( instance classes #26792, array classes #1639) (malloc=5740KB #87822) (mmap: reserved=1189888KB, committed=160048KB) ( Metadata: ) ( reserved=141312KB, committed=139876KB) ( used=135945KB) ( free=3931KB) ( waste=0KB =0.00%) ( Class space:) ( reserved=1048576KB, committed=20172KB) ( used=17864KB) ( free=2308KB) ( waste=0KB =0.00%) - Thread (reserved=696395KB, committed=85455KB) (thread #674) (stack: reserved=692812KB, committed=81872KB) (malloc=2432KB #4046) (arena=1150KB #1347) - Code (reserved=251877KB, committed=105201KB) (malloc=4189KB #11718) (mmap: reserved=247688KB, committed=101012KB) - GC (reserved=230739KB, committed=230739KB) (malloc=32031KB #63631) (mmap: reserved=198708KB, committed=198708KB) - Compiler (reserved=5914KB, committed=5914KB) (malloc=6143KB #3281) (arena=180KB #5) - Internal (reserved=24460KB, committed=24460KB) (malloc=24460KB #13140) - Other (reserved=267034KB, committed=267034KB) (malloc=267034KB #631) - Symbol (reserved=28915KB, committed=28915KB) (malloc=25423KB #330973) (arena=3492KB #1) - Native Memory Tracking (reserved=8433KB, committed=8433KB) (malloc=117KB #1498) (tracking overhead=8316KB) - Arena Chunk (reserved=217KB, committed=217KB) (malloc=217KB) - Logging (reserved=7KB, committed=7KB) (malloc=7KB #266) - Arguments (reserved=19KB, committed=19KB) (malloc=19KB #521) Total: reserved=7168324KB, committed=5380868KB Class (reserved=1195628KB, committed=165788KB) Thread (reserved=696395KB, committed=85455KB) Code (reserved=251877KB, committed=105201KB) GC (reserved=230739KB, committed=230739KB) Compiler (reserved=5914KB, committed=5914KB) Internal (reserved=24460KB, committed=24460KB) Other (reserved=267034KB, committed=267034KB) 

Direct byte buffers Those are the memory segments that are allocated outside the Java heap. Unused buffers are only freed upon GC. Netty for example use them. ● < JDK 11, they are reported in the Internal section ● ≥ JDK 11, they are reported in the Other section Internal (reserved=24460KB, committed=24460KB) Other (reserved=267034KB, committed=267034KB) 

Garbage Collection GC is actually more than only taking care of the garbage. It’s a full blown memory management for the Java Heap, and it requires memory for its internal data structures (E.g. for G1 regions, remembered sets, etc.) On small containers this might be a thing to consider GC (reserved=230739KB, committed=230739KB) 

Threads Threads also appear to take some space Thread (reserved=696395KB, committed=85455KB) (thread #674) (stack: reserved=692812KB, committed=81872KB) (malloc=2432KB #4046) (arena=1150KB #1347) 

Native Memory Tracking 👍 Good insights on the JVM sub-systems 

Native Memory Tracking Good insights on the JVM sub-systems, but Does NMT show everything ? Is NMT data correct ? ⚠ Careful about the overhead! Measure if this is important for you ! 

Huh what virtual, committed, reserved memory? virtual memory : memory management technique that provides an "idealized abstraction of the storage resources that are actually available on a given machine" which "creates the illusion to users of a very large memory". reserved memory : Contiguous chunk of memory from the virtual memory that the program requested to the OS. committed memory : writable subset of reserved memory, might be backed by physical storage 

Native Memory Tracking Basically what NMT show is this 

Native Memory Tracking Basically what NMT show is this 

Native Memory Tracking Basically what NMT show is this 

Huh, what virtual, committed, reserved memory? used heap : amount of memory occupied by live objects and to a certain extent object that are unreachable but not yet collected by the GC committed heap : the size of the writable heap memory where the JVM can write objects. This value seats between -Xms and -Xmx values heap max size : the limit of the heap (-Xmx) #JVM 

Native Memory Tracking Basically what NMT show is : how the JVM subsystems are using the available space 

Native Memory Tracking Good insights on the JVM sub-systems, but Does NMT show everything ? Is NMT data correct ? 

So virtual memory ? 

Virtual memory ? Virtual memory implies memory management. It is a OS feature ● to maximize the utilization of the physical RAM ● to reduce the complexity of handling shared access to physical RAM By providing processes an abstraction of the available memory 

Virtual memory On Linux memory is split in pages (usually 4 KiB) Never used pages remain virtual, that is without physical storage Used pages is called Resident memory 

Virtual memory The numbers shown in NMT are actually about what the JVM asked for. Total: reserved=7168324KB, committed=5380868KB 

Virtual memory Not the real memory usage Total: reserved=7168324KB, committed=5380868KB 

Native Memory Tracking Good insights on the JVM sub-systems, but Does NMT show everything ? Nope Is NMT data correct ? Yes, but not for resident memory usage 

What does it means with JVM flags ? For the Java heap, -Xms / -Xmx ⇒ indication of how much memory heap memory is reserved 

What does it means with JVM flags ? For the Java heap, -Xms / -Xmx ⇒ indication of how much memory heap memory is reserved Also -XX:MaxPermSize, -XX:MaxMetaspaceSize, -Xss, -XX:MaxDirectMemorySize ⇒ indication of how much memory heap memory is/can be reserved These flags do have a big impact on JVM subsystems, as they may trigger or not some behaviors, like : - GC if metaspace is too small - Heap resizing ig Xms ≠ Xmx - … 

💡Memory mapped files They are not reported by Native Memory Tracking, yet They can be accounted in the RSS. 

💡Memory mapped files In Java, using FileChannel.read alone, ⇒ Rely on the native OS read method (in unistd.h) ⇒ Use the OS page cache 

💡Memory mapped files In Java, using FileChannel.read alone, ⇒ Rely on the native OS read method (in unistd.h) ⇒ Use the OS page cache But using FileChannel.map(MapMode, pos, length) ⇒ Rely on the mmap OS method (in sys/mman.h) ⇒ Load the requested content into the addressable space of the process ❌ 

pmap To really deep dive you need to explore the memory mapping ● Via /proc/{pid}/smaps ● Or via pmap [-x|-X] {pid} Address Kbytes RSS Dirty Mode Mapping ... 00007fe51913b000 572180 20280 0 r--s- large-file.tar.xz ... 

How to configure memory requirement preprod prod-asia prod-eu prod-us jvm-service jvm-service jvm-service jvm-service 

How to configure memory requirement Is it possible to extract a formula ? 

How to configure memory requirement Different environments ⇒ different load, different subsystem behavior E.g. preprod : 100 req/s 👉 40 java threads total 👉 mostly liveness endpoints 👉 low versatility in data 👉 low GC activity requirement prod-us : 1000 req/s 👉 200 java threads total 👉 mostly business endpoints 👉 variance in data 👉 higher GC requirements 

How to configure memory requirement Is it possible to extract a formula ? Not that straightforward. Some might point to -XX:*RAMPercentage flags, it sets the Java heap size as a function of the available physical memory. It works. ⚠ -XX:InitialRAMPercentage ⟹ -Xms mem < 96MiB -XX:MinRAMPercentage ⟹ -Xmx mem > 96MiB -XX:MaxRAMPercentage ⟹ -Xmx 

How to configure memory requirement Is it possible to extract a formula ? 1 GiB 4 GiB prod-us preprod MaxRAMPercentage = 85 Java heap Java heap 

How to configure memory requirement Is it possible to extract a formula ? 1 GiB 4 GiB prod-us preprod Java heap MaxRAMPercentage = 85 Java heap 

How to configure memory requirement Is it possible to extract a formula ? 1 GiB 4 GiB prod-us preprod Java heap ≃ 850 MiB Java heap ≃ 3.40 GiB MaxRAMPercentage = 85 

How to configure memory requirement Is it possible to extract a formula ? 1 GiB 4 GiB prod-us preprod Java heap ≃ 850 MiB Java heap ≃ 3.4 GiB MaxRAMPercentage = 85 ~ 150 MiB for every subsystems Maybe OK for quiet workloads ~ 600 MiB for every subsystems Likely not enough for loaded systems ⟹ leads to oomkill 

How to configure memory requirement Traffic, Load are not linear, and do not have linear effects ● MaxRAMPercentage is a linear function of the container available RAM ● Too low MaxRAMPercentage ⟹ waste of space ● Too high MaxRAMPercentage ⟹ risk of oomkills ● Requires to find the sweet spot for all deployments ● Requires to adjust if load changes ● Need to convert back a percentage to raw value 

How to configure memory requirement -XX:*RAMPercentage flags sort of works, Its drawbacks don’t make this quite compelling. ✅ Prefer -Xms / -Xmx 

How to configure memory requirement Let’s have a look at the actual measures RSS 

How to configure memory requirement If Xms and Xmx have the same size, heap is fixed, so focus on “native” memory RSS ● GC internals ● Threads ● Direct memory buffers ● Mapped file buffers ● Metaspace ● Code cache ● … 

How to configure memory requirement It is very hard to predict the actual requirement to for all these. Can we add the values of these zones? Yes but it’s not really maintainable. Don’t mess with until you actually need to! E.g. for each JVM subsystem, you’ll need to understand to predict the actual size. It’s hard, and requires deep knowledge of the JVM. Just don’t ! 

How to configure memory requirement In our experience it’s best to actually retrofit. What does it mean ? Give a larger memory limit to the container, much higher than the max heap size. Heap Container memory limit at 5 GiB 

How to configure memory requirement In our experience it’s best to actually retrofit. What does it mean ? Give a larger memory limit to the container, much higher than the max heap size. 1. Observe the RSS evolution Heap RSS Container memory limit at 5 GiB 

How to configure memory requirement In our experience it’s best to actually retrofit. What does it mean ? Give a larger memory limit to the container, much higher than the max heap size. 1. Observe the RSS evolution 2. If RSS stabilizes after some time Heap RSS RSS stabilizing Container memory limit at 5 GiB 

How to configure memory requirement In our experience it’s best to actually retrofit. What does it mean ? Give a larger memory limit to the container, much higher than the max heap size. 1. Observe the RSS evolution 2. If RSS stabilizes after some time 3. Set the new memory limit with enough leeway (eg 200 MiB) Heap RSS New memory limit With some leeway for RSS increase Container memory limit at 5 GiB 

How to configure memory requirement If the graphs show this RSS less than Heap size 🧐 RSS 

How to configure memory requirement If the graphs show this RSS less than Heap size 🧐 Remember virtual memory! If a page has not been used, then it’s virtual RSS Java heap untouched RSS 

How to configure memory requirement ⚠ If the Java heap is not fully used (as in RSS), the RSS measure to get the max memory utilisation will be wrong To avoid the virtual memory pitfall use -XX:+AlwaysPreTouch All Java heap pages touched RSS 

Memory consumption still look too big 😩 

Memory consumption still look too big 😩 Case in hand with Netty 

Case in hand : Netty Buffers ● Handles pool of DirectByteBuffers (simplified) ● Allocates large chunks and subdivides to satisfy allocations Problem: the more requests to handle the more it may allocate buffers & consume direct memory (native) If not capped ⟹ OOMKill 

Controlling Netty Buffers ● JVM options -XX:MaxDirectMemorySize Hard limit on direct ByteBuffers total size Throws OutOfMemoryError ● Property io.netty.maxDirectMemory to control only Netty buffers? ⇒ No, it’s more complicated 

Controlling Netty Buffers Properties controlling Netty Buffer Pool ThreadLocal Caches! Depends on the number of thread in EventLoops ● io.netty.allocator.cacheTrimInterval ● io.netty.allocator.useCacheForAllThreads ● io.netty.allocation.cacheTrimIntervalMillis ● io.netty.allocator.maxCachedBufferCapacity ● io.netty.allocator.numDirectArenas 

Controlling Netty Buffers ThreadLocal Caches! Depends on the number of thread in EventLoops private static EventLoopGroup getBossGroup(boolean useEpoll) { if (useEpoll) { return new EpollEventLoopGroup(NB_THREADS); } else { return new NioEventLoopGroup(NB_THREADS); } } 

Shaded Netty Buffers Beware of multiple shaded Netty libraries They share the same properties! 

Controlling Netty Buffers 

Case in hand : native allocator Small but steady RSS increase 

Case in hand : native allocator If something doesn’t add up : check the native allocator, but why ? To get memory any program must either ● Call the OS asking for a memory mapping via mmap function ● Call the C standard library malloc function On Linux, standard library = glibc 

Case in hand : native allocator The glibc’s malloc is managing memory via technic called Arena memory management Unfortunately there’s no serviceability tooling around glibc arena management (unless modifying the program to call C API) It may be possible to extrapolate things using a tool like pmap 

Case in hand : native allocator Analyzing memory mapping 00007fe164000000 2736 2736 2736 rw--- [ anon ] 00007fe1642ac000 62800 0 0 ----- [ anon ] Virtual 64 MiB RSS ~ 2.6 MiB 

Case in hand : native allocator Analyzing memory mapping 00007fe164000000 2736 2736 2736 rw--- [ anon ] 00007fe1642ac000 62800 0 0 ----- [ anon ] Virtual 64 MiB x 257 ⟹ RSS ~1.2 GiB RSS ~ 2.6 MiB 

Case in hand : native allocator ● Glibc reacts to CPUs, application threads ● On each access there’s a lock ● Higher number of threads ⟹ higher contention on the arenas ⟹ lead glibc to create more arenas ● There are some tuning options in particular MALLOC_ARENA_MAX, M_MMAP_THRESHOLD, … ⚠ Significant understanding of how glibc’s malloc work, allocation size, etc 

Case in hand : native allocator Better solution, change the application native allocator ● tcmalloc from Google’s gperftools ● jemalloc from Facebook ● minimalloc from Microsoft LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libtcmalloc.so 

Case in hand : native allocator 

Case in hand : native allocator If using tcmalloc or jemalloc, one step away from native allocation profiling. Useful to narrow native memory leak. 

My Container gets re-spawn 

Container Restarted 

Demo minikube + Petclinic 

Quick Fix Increase Probe liveness timeout (either initial delay or interval) livenessProbe: httpGet: path: / port: 8080 initialDelaySeconds: 60 periodSeconds: 10 

JIT Compilation 

Troubleshooting Compile Time Use jstat -compiler to see cumulated compilation time (in s) $ jstat -compiler 1 Compiled Failed Invalid Time FailedType FailedMethod 6002 0 0 101.16 0 

Troubleshooting using JFR Use java -XX:StartFlightRecording jcmd 1 JFR.dump name=1 filename=petclinic.jfr jfr print --events jdk.CompilerStatistics petclinic.jfr 

Troubleshooting using JFR 

Measuring startup time docker run --cpus= -ti spring-petclinic CPUs JVM Startup time (s) Compile time (s) 4 8.402 17.36 2 8.458 10.17 1 15.797 20.22 0.8 20.731 21.71 0.4 41.55 46.51 0.2 86.279 92.93 

C1 vs C2 C1 + C2 C1 only # compiled methods 6,117 5,084 # C1 compiled methods 5,254 5,084 # C2 compiled methods 863 0 Total Time (ms) 21,678 1,234 Total Time in C1 (ms) 2,071 1,234 Total Time in C2 (ms) 19,607 0 

TieredCompilation Interpreter C1 + Profiling C2 Compilation Level 0 3 4 

TieredCompilation queues C2 C1 C2 C2 M1 M2 M3 M4 M5 

TieredCompilation Heuristics Level transitions: ● 0 ➟ 2 ➟ 3 ➟ 4 (C2 Q too long) ● 0 ➟ (3 ➟ 2) ➟ 4 (C1 Q too long, change level in-Q) ● 0 ➟ (3 or 2) ➟ 1 (trivial method or can’t compiled in C2) ● 0 ➟ 4 (can’t compiled in C1) Note: level 3 is 30% slower than level 2 Interpreter C1 + Profiling C2 Comp Level 0 3 4 C1 1 C1 + Limited Profiling 2 

Compiler Settings To only use C1 JIT compiler: -XX:TieredStopAtLevel=1 To adjust C2 compiler threads: -XX:CICompilerCount= 

Measuring startup time docker run --cpus= -ti spring-petclinic CPUs JVM Startup time (s) Compile time (s) JVM Startup time (s) XX:TieredStopAtLevel=1 Compile time (s) 4 8.402 17.36 6.908 (-18%) 1.47 2 8.458 10.17 6.877 (-19%) 1.41 1 15.797 20.22 8.821 (-44%) 1.74 0.8 20.731 21.71 10.857 (-48%) 2.08 0.4 41.55 46.51 22.225 (-47%) 3.67 0.2 86.279 92.93 45.706 (-47%) 6.95 

GC 

Troubleshooting GC Use -Xlog:gc / -XX:+PrintGCDetails 

Troubleshooting GC with JFR/JMC 

Settings properly GC: Metadata Threshold To avoid Full GC for loading more class anre Metaspace resize: Set initial Metaspace size high enough to load all your required classes -XX:MetaspaceSize=512M 

Settings properly GC Use a fixed heap size : -Xms = -Xmx -XX:InitialHeapSize = -XX:MaxHeapSize Heap resize done during Full GC for SerialGC & Parallel GC. G1 is able to resize without FullGC (regions, not the metaspace) 

GC ergonomics: GC selection To verify in log GC (-Xlog:gc): CPU Memory GC < 2 < 2GB Serial ≥ 2 < 2GB Serial < 2 ≥ 2GB Serial ≥ 2 ≥ 2 GB Parallel(

GC ergonomics: # threads selection -XX:ParallelGCThreads= Used for Parallelizing work during STW phases # physical cores ParallelGCThreads ≤ 8 # cores > 8 8 + ⅝ * (# cores - 8) 

GC ergonomics: # threads selection -XX:ConcGCThreads= Used for concurrent work while application is running G1 Shenandoah ZGC Max(ParallelGCThreads + 2 / 4, 1) ¼ # cores ¼ if dynamic or ⅛ # cores 

CPU resource tuning 

CPU Resources shares, quotas ? 

CPU shares Sharing cpu among containers of a node Correspond to Requests for Kubernetes Allow to use all the CPUs if needed sharing with all others containers $ cat /sys/fs/cgroup/cpu.weight 20 $ cat /sys/fs/cgroup/cpu.weight 10 resources: requests: cpu: 500m resources: requests: cpu: 250m 

CPU quotas Fixing limits of CPU used by a container Correspond to Limits to kubernetes resources: limits: cpu: 500m resources: limits: cpu: 250m $ cat /sys/fs/cgroup/cpu.max 50000 100000 $ cat /sys/fs/cgroup/cpu.max 25000 100000 

Shares / Quotas CPU is shared among multiple process. Ill processes could consume all the computing bandwidth. Cgroups help prevent that but require to define boundaries. A 100% A 100% A 100% A 100% C waiting be scheduled B waiting be scheduled 🚦 

Shares / Quotas The lower bound of a CPU request is called shares. A CPU core is divided in 1024 “slices”. A host with 4 CPU will have 4096 shares. 

Shares / Quotas Programs also have the notion of shares. The OS will distributes these computing slices propertionnally. Process asking for 1432 shares (~1.4 CPU) Process asking for 2048 shares (2 CPU) Process asking for 616 shares = 4096 Each are guaranteed to have what they asked for. 

Shares / Quotas Programs also have the notion of shares. The OS will distributes these computing slices propertionnally. Process asking for 1432 shares (~1.4 CPU) Process asking for 2048 shares (2 CPU) Process asking for 616 shares = 4096 Each are guaranteed to have what they asked for. 💡 Upper bounds are not enforced, if there’s CPU available a process can burst 

Shares / Quotas Programs also have the notion of shares. The OS will distributes these computing slices propertionnally. Process asking for 1432 shares (~1.4 CPU) Process asking for 2048 shares (2 CPU) Process asking for 616 shares = 4096 Each are guaranteed to have what they asked for. 💡 Pod schedulers like Kubernetes, use this mechanism to place a pod were enough computation is available 💡 Upper bounds are not enforced, if there’s CPU available a process can burst 

Shares / Quotas Different mechanism used to limit a process. CPU is split in periods of 100 ms (by default) A fraction of a CPU is called millicore, and it’s a thousandth Exemple : 100 * ( 500 / 1000 ) = 50ms resources: limits: cpu: 500m 50 ms per period of 100 ms Period millicores cpu fraction 

Shares / Quotas Now Indeed, it means the limit applies to all accounted cores : 4 CPU ⟹ 4 x 100ms = 400ms resources: limits: cpu: 2500m 250 ms per period of 100 ms 🧐 

Shares / Quotas Shares and quota have nothing to do with a hardware socket resources: limits: cpu: 1 limits: cpu: 1 

Shares / Quotas Shares and quota have nothing to do with a hardware socket resources: limits: cpu: 1 limits: cpu: 1 

Shares / Quotas If the process reaches its limit, it will get throttled ie it will have to wait for the next period. Eg s process can consume a 200ms budget on ● 2 cores with 100ms on each ● 8 cores with 25ms on each 

CPU Throttling When you reach limit with CPU quotas, throttling happens Throttling ⟹ STW pauses Monitor throttling: Cgroup v1: /sys/fs/cgroup/cpu,cpuacct//cpu.stat Cgroup v2: /sys/fs/cgroup/cpu.stat ● nr_periods – number of periods that any thread in the cgroup was runnable ● nr_throttled – number of runnable periods in which the application used its entire quota and was throttled ● throttled_time – sum total amount of time individual threads within the cgroup were throttled 

CPU Throttling with JFR JFR Container event jdk.ContainerCPUThrottling 

availableProcessors ergonomics Setting CPU shares/quotas have a direct impact on Runtime.availableProcessors() API Shares Quotas Period availableProcessors() 4096 -1 100 000 4 (Shares / 1024) 1024 300 000 100 000 3 (Quotas / Period) 

availableProcessors ergonomics Runtime.availableProcessors() API is used to : ● size some concurrent structures ● ForkJoinPool, used for Parallel Streams, CompletableFuture, … 

Tuning CPU Trade-off cpu needs for startup time VS request time ● Adjust CPU shares / CPU quotas ● Adjust liveness timeout ● Use readiness / startup probes 

Conclusion 

Memory ● JVM memory is not only Java heap ● Native parts are less known, and difficult to monitor and estimate ● Yet they are important moving part to account to avoid OOMKills ● Bonus revise virtual memory 

Startup ● Containers with <2 cpus are an constraint environment for JVM ● Need to keep in mind that JVM subsystems like JIT or GC need to be adjusted for requirements ● To be aware of these subsystems helps to find the balance between resources and requirements of your application 

References 

References Using Jdk Flight Recorder and Jdk Mission Control MaxRAMPercentage is not what I whished for Off-Heap reconnaissance Startup, Containers and TieredCompilation Hotspot JVM performance tuning guidelines Application Dynamic Class Data Sharing in HotSpot JVM Jdk18 G1 Parallel GC Changes Unthrottled fixing cpu limits in the cloud Best Practices Java single-core containers Containerize your Java applications