SPECjbb MultiJVM - Java Performance

Moving on from SPECCPU, we shift over to SPECjbb2015. SPECjbb is a from ground-up developed benchmark that aims to cover both Java performance and server-like workloads, from the SPEC website:

“The SPECjbb2015 benchmark is based on the usage model of a worldwide supermarket company with an IT infrastructure that handles a mix of point-of-sale requests, online purchases, and data-mining operations. It exercises Java 7 and higher features, using the latest data formats (XML), communication using compression, and secure messaging.

Performance metrics are provided for both pure throughput and critical throughput under service-level agreements (SLAs), with response times ranging from 10 to 100 milliseconds.”

The important thing to note here is that the workload is of a transactional nature that mostly works on the data-plane, between different Java virtual machines, and thus threads.

We’re using the MultiJVM test method where as all the benchmark components, meaning controller, server and client virtual machines are running on the same physical machine.

The JVM runtime we’re using is OpenJDK 15 on both x86 and Arm platforms, although not exactly the same sub-version, but closest we could get:

EPYC & Xeon systems:

openjdk 15 2020-09-15
OpenJDK Runtime Environment (build 15+36-Ubuntu-1)
OpenJDK 64-Bit Server VM (build 15+36-Ubuntu-1, mixed mode, sharing)

Altra system:

openjdk 15.0.1 2020-10-20
OpenJDK Runtime Environment 20.9 (build 15.0.1+9)
OpenJDK 64-Bit Server VM 20.9 (build 15.0.1+9, mixed mode, sharing)

Furthermore, we’re configuring SPECjbb’s runtime settings with the following configurables:

SPEC_OPTS_C="-Dspecjbb.group.count=$GROUP_COUNT -Dspecjbb.txi.pergroup.count=$TI_JVM_COUNT -Dspecjbb.forkjoin.workers=N -Dspecjbb.forkjoin.workers.Tier1=N -Dspecjbb.forkjoin.workers.Tier2=1 -Dspecjbb.forkjoin.workers.Tier3=16"

Where N=160 for 2S Altra test runs, N=80 for 1S Altra test runs, N=112 for 2S Xeon 8280, N=56 for 1S Xeon 8280, and N=128 for 2S and 1S on the EPYC system. The 75F3 system had the worker count reduced to 64 and 32 for 2S/1S runs.

The Xeon 8380 was running at N=140 for 2S Xeon 8380 and N=70 for 1S - the benchmark had been erroring out at higher thread counts.

In terms of JVM options, we’re limiting ourselves to bare-bone options to keep things simple and straightforward:

EPYC & Altra systems:

JAVA_OPTS_C="-server -Xms2g -Xmx2g -Xmn1536m -XX:+UseParallelGC "
JAVA_OPTS_TI="-server -Xms2g -Xmx2g -Xmn1536m -XX:+UseParallelGC"
JAVA_OPTS_BE="-server -Xms48g -Xmx48g -Xmn42g -XX:+UseParallelGC -XX:+AlwaysPreTouch"

Xeon Cascade Lake systems:

JAVA_OPTS_C="-server -Xms2g -Xmx2g -Xmn1536m -XX:+UseParallelGC"
JAVA_OPTS_TI="-server -Xms2g -Xmx2g -Xmn1536m -XX:+UseParallelGC"
JAVA_OPTS_BE="-server -Xms172g -Xmx172g -Xmn156g -XX:+UseParallelGC -XX:+AlwaysPreTouch"

Xeon Ice Lake systems (SNC1):

JAVA_OPTS_C="-server -Xms2g -Xmx2g -Xmn1536m -XX:+UseParallelGC"
JAVA_OPTS_TI="-server -Xms2g -Xmx2g -Xmn1536m -XX:+UseParallelGC"
JAVA_OPTS_BE="-server -Xms192g -Xmx192g -Xmn168g -XX:+UseParallelGC -XX:+AlwaysPreTouch"

Xeon Ice Lake systems (SNC2):

JAVA_OPTS_C="-server -Xms2g -Xmx2g -Xmn1536m -XX:+UseParallelGC"
JAVA_OPTS_TI="-server -Xms2g -Xmx2g -Xmn1536m -XX:+UseParallelGC"
JAVA_OPTS_BE="-server -Xms96g -Xmx96g -Xmn84g -XX:+UseParallelGC -XX:+AlwaysPreTouch"

The reason the Xeon CLX system is running a larger back-end heap is because we’re running a single NUMA node per socket, while for the Altra and EPYC we’re running four NUMA nodes per socket for maximised throughput, meaning for the 2S figures we have 8 backends running for the Altra and EPYC and 2 for the Xeon, and naturally half of those numbers for the 1S benchmarks.

For the Ice Lake system, I ran both SNC1 (one NUMA node) as SNC2 (two nodes), with the corresponding scaling in the back-end memory allocation.

The back-ends and transaction injectors are affinitised to their local NUMA node with numactl –cpunodebind and –membind, while the controller is called with –interleave=all.

The max-jOPS and critical-jOPS result figures are defined as follows:

"The max-jOPS is the last successful injection rate before the first failing injection rate where the reattempt also fails. For example, if during the RT-curve phase the injection rate of 80000 passes, but the next injection rate of 90000 fails on two successive attempts, then the max-jOPS would be 80000."

"The overall critical-jOPS is computed by taking the geomean of the individual critical-jOPS computed at these five SLA points, namely:

      • Critical-jOPSoverall = Geo-mean of (critical-jOPS@ 10ms, 25ms, 50ms, 75ms and 100ms response time SLAs)

During the RT curve building phase the Transaction Injector measures the 99th percentile response times at each step level for all the requests (see section 9) that are considered in the metrics computations. It then computes the Critical-jOPS for each of the above five SLA points using the following formula:
(first * nOver + last * nUnder) / (nOver + nUnder) "


That’s a lot of technicalities to explain an admittedly complex benchmark, but the gist of it is that max-jOPS represents the maximum transaction throughput of a system until further requests fail, and critical-jOPS is an aggregate geomean transaction throughput within several levels of guaranteed response times, essentially different levels of quality of service.

Beyond the result figures, the benchmark keeps detailed track of timings of responses and tracks a few important statistical data-points across a response-time curve, as follows:


2S Xeon 8380 THP Enabled


2S Xeon 8280 THP Enabled 

Comparing the Xeon 8380 to the Xeon 8280, what’s to be immediately noted is the much-improved maximum throughput figure of the new part, scaling at +64% compared to its predecessor. We’re seeing that the load slope where the 99th percentile SLA figures rises comes in at a relative earlier point, and the corresponding critical-jOPS point lands in relatively earlier than the Xeon 8280.


2S EPYC 7763 THP Enabled


2S Altra Q80-33 THP Enabled

I included the AMD EPYC 7763 and Altra graphs for context.

SPECjbb2015-MultiJVM max-jOPS

 As commented in the response curve analysis, the new Xeon 8380 sees huge leaps in the max-jOPS metrics, vastly outperforming the Xeon 8280 and landing in a very favourable competitive positioning compared to the AMD parts.

My theory here is that because of the good per-core performance of the Intel design, along with the monolithic mesh architecture, while Intel doesn’t quite catch up with AMD, it performs very well with relatively significantly fewer cores.

SPECjbb2015-MultiJVM critical-jOPS

 The critical-jOPS metric however wasn’t quite as positive for the new Xeon 8380. Although the chip is showing increases in performance, there’s not as strong as the max-jOPS measurements. At first I had measured the SNC off mode of the platform, similar to the 8280 numbers we have (Our ASRock test bed doesn’t expose SNC options in the BIOS), however these results were extremely meagre as they barely differentiated to the 8280. Running the system in SNC2 mode actually improved the critical jOPS figure more significantly, whilst only marginally affecting the max-jOPS metric.

What’s really odd about the results though is that this larger increase only happens in the 2S test figures, with the 1S being unfavourable to the new Ice Lake part, losing to the 8280 in both modes. I had repeated these numbers several times to be sure they’re repeatable, and they were indeed so – as odd as that is. The 1S reduction in the critical-jOPS could be explained through the larger mesh size and larger core count of the 8380, and we did see slight regressions in core-to-core latencies. If the mesh intersection bandwidth did not increase with its size, that also could be a culprit of these figures, as the workload is hammering core-to-core transactions as well as the L3 cache of the chip.

Why the 2S figures see a bigger advantage of migrating to SNC2 could be a result of how on-chip traffic is routed, as well as the traffic flows through the UPI link blocks of the chip – at least that would be my working hypothesis.

Intel had disclosed a +62% figure for a “Java Throughput under SLA” workload they wouldn’t specify, and this does track well with our max-jOPS results. While the critical-jOPS increases seem a bit disappointing generationally, how it translates to the real world in contrast to the max-jOPS figure depends on how strict one’s SLA metrics are.

SPEC - Per-Core Performance under Load Compiling LLVM, NAMD Performance
Comments Locked

169 Comments

View All Comments

  • lmcd - Tuesday, April 6, 2021 - link

    Linking semiaccurate like it's accurate, the jokes write themselves.
  • arashi - Tuesday, April 6, 2021 - link

    Still more accurate than the embarrassment called #silicongang.
  • schujj07 - Tuesday, April 6, 2021 - link

    As an actual administrator in a datacenter your statement about those advantages is bogus.
  • Hifihedgehog - Tuesday, April 6, 2021 - link

    > AMD, as we know even from consumer products isn't that amazing when it comes to drivers, BIOS quality and fixing bugs, whereas Intel is much more reliable in this regard.

    What drivel even is this? Have you actually worked in the industry? Clearly, you have not. I have already seen machine learning nodes move to Epyc, my web host has since moved to Epyc, and even a lot of recommendations for home lab equipment (see ServeTheHome) has since been moving heavily towards AMD. You have no clue. So go eat a pound of sand. At least it will put out better crap than Intel’s 10nm.
  • amootpoint - Wednesday, April 7, 2021 - link

    If your ML has moved to AMD, you are already burning a lot of money ... good luck.

    AI is where AMD is lagging so much compared to Intel, that it doesn’t even make sense.
  • schujj07 - Wednesday, April 7, 2021 - link

    You obviously don't know what the term "Machine Learning Node" actually means. That doesn't mean the accelerators for machine learning are FirePro or Epyc, just the server that houses them are running Epyc.
  • amootpoint - Thursday, April 8, 2021 - link

    You clearly sounds like an arrogant guy, with full on personal attacks. No point in further discussion.
  • schujj07 - Thursday, April 8, 2021 - link

    Pot calling kettle black.
  • duploxxx - Wednesday, April 7, 2021 - link

    This is a release of server chips, which are distributed through OEM mainly with their specific drivers and BIOS releases close design with AMD.... Do you honestly believe that you get instable BIOS, drivers for Server releases? It's not a consumer moboproduct for 50-100-150$ that goes on sale for the masses with generic subset of BIOS that needs to look fancy, has oc potential, looks and feel, fan mngmnt etc....

    second, price, upgradeability is still in favor of the AMD product, quality and support is delivered by the same OEM that ships both intel and AMD systems... and performance, well we know that answer already.. which only leaves retared ICT members that are aging and still believe in some mythical vendors... well i hope they still like all the spectre and meltdown patches and feel evry confident in there infinit support of a dominant monopoly and like to pay 10-15k$ for a server cpu to allow a bit more Ram support.
  • DannyH246 - Tuesday, April 6, 2021 - link

    Oh and also don’t forget the usual Intel fused off features just because...Compare this to AMD where you get all features in all SKU’s. Anyone who recommends this crap is simply an Intel Fanboi.

Log in

Don't have an account? Sign up now