Cavium ThunderX2 Review and Benchmarks a Real Arm Server Option

  • Thread starter Patrick Kennedy
  • Start date
Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.

Patrick

Administrator
Staff member
Dec 21, 2010
12,511
5,792
113
Thanks guys. That took a long time to get done. We will have more follow-up but I wanted to get something out this week.

I also spoke at the GA event. Will post video when/if I get it.
 
  • Like
Reactions: eva2000 and K D

Patrick

Administrator
Staff member
Dec 21, 2010
12,511
5,792
113
This has been a pretty big article for us. I am surprised nobody has posted it on Hacker News yet. Good feedback thus far.
 
  • Like
Reactions: eva2000

zir_blazer

Active Member
Dec 5, 2016
355
128
43
Whoa, ARM standing in brute performance against enterprise x86. Didn't saw it coming, was expecting POWER to do that.

Note that there should be some typos on the parts table.
CN9978-1800LG4077-Y21-G
CN9970-1800LG4077-Y21-G
Both of these appear with both a 8 MC and 4 MC but under the same ordering code. I think some other models also have that issue. Is the same model supposed to do both roles or is a typo and the ordering code should be different?
 

mstone

Active Member
Mar 11, 2015
505
118
43
46
Thanks @Patrick great write up. Would be great to see OpenSSL ECDSA 256bit performance too :)

Code:
openssl speed -multi $(nproc) rsa4096 rsa2048 ecdsap256
I thought at one point there was agreement that we would get AES (+NI) results (as the RSA results are frankly irrelevant--it's not used for bulk encryption, and is rarely a bottleneck; AES is what you're typically going to be relying on for disk encryption, network encryption, etc., and is a much more interesting number).
 

Patrick

Administrator
Staff member
Dec 21, 2010
12,511
5,792
113
@zir_blazer that is Cavium's table.

@eva2000 - call this not scientific but the ecdsap256 part of that is (sign/s and verify/s)
2P CN9980 1406779.2 582276.0
2P 6148 941411.0 429894.5
2P 7601 643925.0 612235.4

Running on straight Ubuntu 18.04 LTS w/ OpenSSL 1.1.0g no tuning at all. We do have a follow-up piece planned so if there is anything else you want to see let me know.
 
  • Like
Reactions: eva2000

eva2000

Active Member
Apr 15, 2013
244
49
28
Brisbane, Australia
centminmod.com
Cheers thanks for that - just formatting it for ease of viewing :)
Code:
2P  CN9980  1406779.2  582276.0
2P  6148    941411.0   429894.5
2P  7601    643925.0   612235.4
That CN9980 is a beast at ECDSA signs/s !

Something like a Nginx HTTP/2 HTTPS reverse proxy using ECDSA 256bit vs RSA2048bit SSL certificates would be interesting to see.

Or Redis stunnel configuration using ECDSA 256bit SSL certs on both ends. Been testing Redis stunnel optimisations and found that ECDSA 256bit SSL certs shine nicely compared to traditional RSA 2048bit stunnel setups especially with OpenSSL 1.1.1 and TLSv1.3 centminmod/centminmod-stunnel

upload_2018-5-12_3-42-13.png

Though Redis stunnel would be more single threaded work loads.
 
Last edited:

mstone

Active Member
Mar 11, 2015
505
118
43
46
Something like a Nginx HTTP/2 HTTPS reverse proxy using ECDSA 256bit vs RSA2048bit SSL certificates would be interesting to see.
No, it would seem like a waste of resources in a mis-spec'd box to use something with this many cores & pcie channels & an expensive NUMA interconnect to fill a role that scales horizontally with no licensing costs.
 

Saurabh Bagh

New Member
Feb 6, 2019
2
0
1
I saw the with Linkpack you have shown Cavium Thunder X2 to reach the benchmark of 1000 GFlops without the ARM compiler in your website.
I have been trying to do the benchmarking of Thunder X2 and I can't even reach the 100 GFlops. I have tried a lot of combinations and still no success :(
I would really appreciate if someone would help me out on that , as it is my university project
In case you want, I can share my HPL.dat and Config file