admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
GSB benchmarking
Im intending to make an option to benchmark GSB(s)
ie 1 GSB takes 30 minutes on i9_7900
ie 10 GSB takes 120 minutes on i9_7900
//The figures are a guess.
We will need a standard setup, and the random seed needs to be fixed to results are the same every time.
Let me know if this if interest.
The purpose is to test what hardware is fastest and or best bang for buck
|
|
|
JasonT
Junior Member
Posts: 61
Registered: 6-6-2018
Member Is Offline
Mood: No Mood
|
|
Great idea. Happy to perform a test or two and provide results back.
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
I will get you a copy. What cpu / ram do you have?
Ive learnt some things, but this takes a lot of cpu time.
reason is im not yet sure if a 15 generation test is as valid as a 40 generation.
SO far I suspect having the cpu a bit lower than max is best.
will know more in a few days. A lot of programing time could go into this project and im not sure where to stop. Depends also if its free or not. It
certainly affects us financially as we get a lot better idea of what h/w to buy.
The second stage is going to be WF benchmark.
ie how many wf per worker, and how many wf per machine, and single or multi threaded
|
|
|
Carl
Member
Posts: 342
Registered: 10-5-2017
Member Is Offline
Mood: No Mood
|
|
Hi Peter,
Please send me a copy, so I can run it on my machine.
i7-7700, 32 GB RAM
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
I shall get it tomorrow else monday
I need to prepare all the files...
here is a screen shot.
The new version I just started has ram and cpu measurements.
the old one didnt but you can see 8 gsb was narrowly the best. (40 logical processors)
|
|
|
emsjoflo
Junior Member
Posts: 39
Registered: 7-7-2017
Member Is Offline
Mood: No Mood
|
|
My observations from GSB Benchmark
I purchased GSB Benchmark to help me decide whether to retire my old server. And to help me decide how to configure a newer server to get the best
performance per $. GSB's performance is not linear with other benchmarks (ie. a higher passmark or Cinebench score does not guarantee a faster GSB
execution). My theory is that GSB uses multiple cores fairly efficiently but benefits from a high clock speed because of a few single-thread
functions.
And since the default settings on the resource manager seem to keep the processor between 80 and 90%, that is what I'm looking at.
My understanding of Peter's GSB Benchmark is that it automates running multiple instances of GSB standalone and keeps track of execution time, ram and
CPU utilization. It takes about 6 hours for me to run a full test on my machine but it generates some interesting data. If others want to share
their GSB Benchmark data, I would be willing to compile data into a spreadsheet.
I look for several things:
1. How can I optimize software settings to get the most performance out of the machine? From what I have seen, GSB is ram hungry so larger cache
settings seem to allow the processor to run faster per instance.
2. What are the bottlenecks for the PC being tested? For the two older PC's I have tested, the CPU reaches 80% utilization before it maxes out the
ram. In both cases, after the CPU reaches 92% utilization the average execution time gets worse with additional instances. With the dual X5670, 8
instances gave the lowest Avg execution time at 2.88s with 71G ram and 92% CPU. But in actual practice, RM doesn't allow more than 4 instances to run
(just over 3s execution time) -- probably because the small amount of overhead required by RM and GSBmanager push the CPU % up a bit more. More than
96G of ram wouldn't make the X5670 faster because the CPU is maxed out at just over half. The dual E5-2670 server actually saturates the CPU with
less instances and gets the best score with 6 instances at 2.27s with 49G and 94% CPU. But real world, RM might only allow 4 instances (2.5s). More
ram probably wouldn't benefit the 2670 either.
And based on the info Peter shared with me, it seems that ram speed isn't a bottleneck -- DDR4 ram gave no noticeable benefit over DDR3.
Peter's dual 2690v2 didn't reach 90% until 14 instances and 147G of ram. So the newer, faster processors benefit from at least 196G of ram. His
2699v3 with 64G only got up to 60%CPU with 5 instances before using up the ram.
3. What is the best value for a GSB PC? Peter thinks a dual E5-2680v2 server with 256gb of ram is the best value and he is probably right. I think
a 2667 might be pretty good too. In the test results he released to me, the dual 2680v2 performed better than i9's and a single 2699v3. An i7 was
about 2.5 times slower, dual 2620v2 about 2 times slower, my old dual x5670 was about 1.8 times slower and the dual 2670 about 1.4 times slower, I9
about 1.2 times slower (compared to the best times of the dual 2680v2). Peter wasn't impressed with his last test of the AMD threadripper but I'm not
aware of any GSB tests on the newer AMD chips.
the best prices I've found on Ebay(US) are as follows:
$700 2x E5-2690 128gb SSD
$1081 2x E5-2687Wv2 192gb SSD
$1164 2x E5-2680v2 256gb SSD
$1200 2x E5-2667v2 256gb SSD
$1355 2x E5-2690v2 256gb SSD
$1600 2x E5-2697v2 256gb SSD
$2600 2x E5-2690v3 256gb SSD
$2700 2x E5-2687Wv3 256gb SSD
$2800 2x E5-2699v3 256gb SSD
And the best price I've found for colocation in the midwest US (same timezone as Chicago) is $79/m for a 1U dual-Xeon server with redundant power
supplies. Steadfast is considerably more.
The best prices I've found for windows dedicated servers are as follows:
$99 2x2670 96gb + windows license fee
$104 i7-6700 64gb
$114 2x2660 128gb (+20/m windows license fee or install own OS w KVM)
$134 2x2670v2 128gb (+20/m windows license fee or install own OS w KVM)
$210 2x2667v2 192gb
$280 2x2695v2 256gb
$282 2x2678v3 256gb
Renting from Peter is another good option -- I've spent over a week researching and about 3 days configuring my rented server (getting software
installed and licensed, downloading updates and drivers, configuring power settings, optimizing ram and swap files, etc.)
Anyway, I decided to rent at least one server for now to test, the rental isn't much more than colo fees. And if I need more power, I might just
rent another cheap server.
 
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
thanks for your research. I hope to have results of a GSB user with dual xeon 2690v3 soon.
There is also a new build on BM out today via auto update. The old one expired.
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
I may also do rent of 50% of a server too. Typically a dual xeon with 256gb of ram is happy to run about 10 workers.
That way you will get 5 workers for $5 a day
Ive ordered 5 more servers in the last month, though 2 were for long term hire. This will help the free cloud a bit, though it depends greatly if I am
using all my cloud or not.
|
|
|
admin
Super Administrator
Posts: 5060
Registered: 7-4-2017
Member Is Offline
Mood: No Mood
|
|
The new installer of GSB Benchmark is in the private forum, here
https://trademaid.info/forum/viewthread.php?tid=250#pid4410
|
|
|
![[*]](images/davis/default_icon.gif){kind=icon}
