I recently purchased some new to me r620 servers for a cluster. Mostly they will be doing heavy database transactions, but generally they will have Hyper-V vms doing a variety of work. It was during the database work that I started realizing that the servers were performing much worse than my old r610. Since then I've swapped out controllers, nics, and drives in search of performance comparable to other diskmark tests on similar systems posted online. Mostly my random single threaded performance seems to be horrible. Changing the bios to Performance helped a lot, but I'm still running slow. Enabling/Disabling read, write, and disk cache changes behavior, but does not alter performance radically either way. Every update is applied, and using no read ahead/write back/disk cache enabled for tests (best results). Am I missing something, could my CPU really be that much of a single thread bottleneck, or are my results normal? Thanks for any advice!
System:
R620
Windows Server 2019 Core with Hyper-V – Server 2019 and Ubuntu 18.04 guests
Dual E5-2650v2
128GB (16x8GB PC3L-12800R)
H710p mini mono
5x Intel D3-S4610 960GB SSDs in Raid 5
Intel X540 NIC
Using CrystalMark 3 – 9/4GB:
My system
Read / Write
Seq: 1018 / 1637
512K: 743 / 1158
4K: 19 / 23
4k QD32: 204 / 75
Comparison system – https://www.brentozar.com/archive/2013/08/load-testing-solid-state-drives-raid/
Read / Write
Seq: 1855 / 1912
512K: 1480 / 1419
4K: 34 / 51
4k QD32: 651 / 88
Using CrystalMark 6 – 2/100mb:
my system
Read / Write
Seq Q32T1: 3022 / 3461
4k Q8T8: 335 / 290
4K Q32T1: 210 / 195
4K Q1T1: 32 / 30
Comparison system – https://www.youtube.com/watch?v=i-eCmE5itzM
Read / Write
Seq Q32T1: 554 / 264
4k Q8T8: 314 / 259
4K Q32T1: 316 / 261
4K Q1T1: 33 / 115
Using CrystalMark 6 – 5/1GB:
My system
Read / Write
Seq Q32T1: 2619 / 1957
4k Q8T8: 306 / 132
4K Q32T1: 212 / 116
4K Q1T1: 25 / 27
Comparison system – R610, Hyper-V Core 2012R2 -2008R2 Guests – Dual X5670, 128 GB 1600mhz ram, 4x Samsung 860 Pro 1TB raid 5, h700
Read / Write
Seq Q32T1: 754 / 685
4k Q8T8: 305 / 69
4K Q32T1: 262 / 69
4K Q1T1: 32 / 38
Here are some real world numbers compared to my old R610 system
Export same database table from a local mariadb to a single R620 Mariadb Galera cluster node
R610 – 1.7 million recs/min
R620 – 1.16 million recs/min
Copy folder with thousands of small files from VM to Host
R610 – 23 seconds
R620 – 2 min 40 seconds
Alternatively, large file copies show good performance with R620 beating R610 by about 35%.
Best Answer
Server 2019 is the problem after all. I've tried tweaking every setting, changing every piece of hardware, and updating everything to current as of May 2019. In the end the system performed well out of the box with Server 2016.