D-1540 or E5v3 - Torn and need more info for usecase

[Ramos]

Getting in a lot of doubt here. Been looking back and forth at the XU4's CPU benchmarks of various kinds and compared it to the J1900, but it's very hard to find some that compare apples to apples.

Closest I got was Sysbench on 2000 primes in 1.7 sec on the XU4 and 10k primes in 11 sec on the J1900.

Does anyone have an XU4/J1900 and can run some tests for me? (that has been run for the other device already)

Specifically I am looking for J1900 marks for the tests in the XU4 column,
CPU Boards - performance comparison - Open Glider Network Project

I know the XU4 is pretty darn good at everything but CPU marks and even there its very decent, but the J1900 series has the same kinda power requirements (almost) and lots better flexibility (pci-e card slots, ram can vary) and the price is nearly the same.

 

[Ramos]

Just for closure, thanks for advice here guys. I ended up with two dual E5-2670s from Natex. The primary factor was that they are almost giving them away and that power won't be worse than the 130W rig I already got. I plan on selling one and get some of the shipping to Denmark covered that way. Shipping was the same either way so...

And the 128GB rig was perfect for what I actually need (eventually when maxed out at the same time):
- 8 nodes, worst case ( on ESXi )
- 1 x Win10 Pro deskoping.
- 1 x Win2k16 Server for all the server services and maybe Hyper-V, though I plan on using ESXi.
Virtual Cluster (4-6 nodes),
- 2 x Master nodes (only for failover testing)
- 4 x Data/Slave/Worker nodes

All can now have 16 GB RAM and 2c each, though I will probably switch cores around alot depending on cluster test.

Cluster will be testbench for (in 2016 at least):
- PXE deployment of nodes automatically.
- MemSQL,
- IBM Platform Conductor for Spark
- MongoDB shardtesting
- Distributed / Parallel computing in general (M-L, image learning, etc)
- Hadoop (various distros, BigSQL/Impala via BigInsights/CDM at least)
- Perhaps an AP NoSQL db like Cassandra as well.
- "Oak table" comparison testing of claims.

The power problem will have to be settled as good as possible with a 80+ Plat/Titan.
Basically I am copying @gigatexal 's build as much as I can on the one I keep.

As for a minicluster on Odroids, I might test an XU4 as datanode vs a J1900, but still maybe. There wont be a new XU5 till 2017, Odroid confirmed that on a forum post of theirs sadly. I might just wait till then as the J1900 are supernice but dated (2014) and there might be a newer one with more power then. Said the guy who just bought a 2011 rig ...

 

[gigatexal]

Ramos you're testing memsql? that's pretty cool. I'm interested in that as well that and the other nosql solutions you listed. You should create a thread about what you find. I'm no @dba but it's what I do for a living. I'm also glad the 2670 setup worked out for you!

I've a seasonic 660XP platinum PSU that powers my get-up like a champ, even with a power hungry 390X gpu

 

[Ramos]

Will do report results but it won't be before like June or so. Sounds good with the PSU, I was aiming for an EVGA P2 Supernova 650W cause its 25 euro cheaper over here and I have good prior vibes with those.

Is the 390X the choice because of the OpenCL or because the AMD cards rocked bitcoin mining so maybe they are the faster ones for GPU programming when we talk gamer cards?
(I cannot afford the Tesla cards for my own rig)

 

[gigatexal]

i'm about to throw fedora 23 on my current dual 2670s and tinker with python and therano with my 390x. I got it for gaming but I do less of that and more programming of late. From what I can tell CUDA and Tesla cards are the defacto standards but it's hard to pass up the relativley cheap sheer performance from AMD

 

[T_Minus]

memsql is old old right? never took off, lots of hype.. at-least I Think I'm recalling that right