Single Socket Server for AMD EPYC is a great choice for your RDBMS, why?
This test is actually has been done last year, but we just able to finalize and show this report to public just recently since pandemic is also delivery more business to us. It make us very happy busy…
We were doing Zen 2 (Rome) tests for TPC-B and TPC-C on single and dual sockets, and also run Zen 3 (Milan) test TPC-B and TPC-C on single and dual sockets as well.
A short report for you who are very busy: Use AMD EPYC Processors for all your needs, and use the single socket one for your very high performance RDBMS such as PostgreSQL, Oracle, etc.
Last year is a tough year indeed, in the middle of Covid19 fight we also trying to test our appetite of hungry of computing power and has been lucky enough to test the Zen 3 processors as in replacement of our Zen 2 processors. In Zen 2 processors, we have been testing its performance using pgbench for TPC-B and HammerDb for TPC-C PostgreSQL performance, and it results single socket is more efficient than dual socket.
Some people may think, if single processor has delivery such great performance, why not we get both? this is a general question, the answer is it depends.
Why, it depends on what?
It depend on how you wanted to use the server, it may be a great investment to have dual socket server since the price will be more efficient. Apparently dual socket server of AMD EPYC cost less than 50% expensive compare to single socket. If the usage of the great power server is not really relate to the coordination between core, it will perhaps having a liniear improvement from single socket compare to dual socket one. Otherwise it won’t.
In a dual socket environment of Zen 2, the core basis performance decreased almost 50%, it means dual socket environment only improve total server performance no bigger than 3%. It apparently Zen 2 processors has been perform worse in dual socket environment, due to the lack of great communication in between sockets.
So, by those fact we are recommending user to use single socket instead of dual for RDBMS usage. You may check the complete report by download from the link below.
How is the difference between Zen 2 Architecture compare to Zen 3?
Zen 3 has proven deliver great improvement, as core per core basis it increase almost 30%. We use the same server, replacing the processor in the same socket and a little bit bios update for running from Zen 2 processor into Zen 3 one. A simple replacement method deliver tremendous performance!
Zen 3 has a better configuration of core dies in the chiplet. This better architecture perhaps contributes better performance. In a single processor of Zen 3, there are some chiplets which each contains of 8 cores and share the L3 cache altogether. And 8 chiplet forms together in a single die become the Processor. 8x8 is 64 real core, or 128 Hyperthreads.
In every chiplet which contains 8 cores, shares L3 caches, and having each dedicated L2 caches per core, communicated and coordinated by a technology called Infinity Fabric which enable a great communication between cores and clusters of chiplets in a synergy to crunch loads every single clock.
Based on our comprehensive test, with a cautious tuning on the Database (PostgreSQL of course) and Operating System, we found and prove that Infinity Fabric implemented in AMD SoC Processors has deliver a great performance especially on the coordinated and synergized system such as RDBMS.
Dual socket Zen 3 is not very well performs while still better than Zen 2. If Zen 2 only contribute increasing performance for about 3%. Zen 3 increases a bit by 18% higher from single socket into dual socket. Both of the test (Zen 2 and Zen 3) shown that dual socket is not efficient to be used in high performance database servers.
Due to the great Infinity Fabrics, EPYC processors prevail any other competitors in this case the prominent contender is Intel processors, single socket of EPYC machine has the power better than dual processors Xeon Machine (Intel) but not for dual socket EPYC.
Dual socket EPYC would perhaps perform better in another scenarion and environment, scope of this discussion of benchmarking EPYC processors in RDBMS OLTP business process test.
For anyone who wanted to procure hardware for your system, you may take a look our elaborated test report in this link: http://equnix.asia/media/benchmark/AMDRomeMilan.pdf
Equnix is not affilated with AMD or Intel brand mentioned here, Equnix stays neutral and run the test with integrity, describe the result and fact as is. For more information you may visit and discuss with us:
Equnix is an IT Solution Provider in Indonesia, we are serving our corporate clients ranging from Bank Industry, Telecommunications, Retails and Governments. We are striving to deliver alternative solution based on Open Source and our state-of-the-art Software Development. We believe that Open Source is a real miracle that given to us have our freedom, our Independence and our Souvereignity.