Platform Threads versus Virtual Threads in SPECjbb2015 : An Empirical Study on OpenJDK 25

Sammanfattning

Virtual threads make the thread-per-request style cheap by unmounting a thread from its carrier when it blocks, but their cost and benefit under asustained server workload are not well understood. This thesis studies them in the SPECjbb2015 benchmark by implementing the Transaction Injector with virtual threads, so the platform-threadand virtual-thread implementations drive the same backend on an unmodified JVM (JDK25). The two are compared under CPU-pinned and unpinned configurations. The results show lower peak throughput for the virtual-thread implementation in both configurations: peak hbIR is 14-17% lower and max-jOPS about 10-12% lower. The latency-bounded critical-jOPS differs between the two CPU configurations: it is 2.6% lower in the unpinnedconfiguration but about 10% higher in the pinned configuration.Bounding the virtual-thread concurrency raises the throughput metrics byabout 16-20% while changing critical-jOPS by no more than 1%. The virtual-thread implementation also uses more CPU, retains more live heap, spends more time in garbage collection, and produces more operating-system level context switches.