𝕏

Benchmarks of beta

Message boards : Number crunching : Benchmarks of beta
Message board moderation

To post messages, you must log in.

AuthorMessage
[VENETO] boboviz

Send message
Joined: 9 Apr 15
Posts: 171
Credit: 1,371,098
RAC: 1,159
Message 838 - Posted: 6 Apr 2016, 9:54:12 UTC

I see a strange behaviour of beta app on my hosts.

1) AMD FX 6300 (2855 float/10898 integer), from 7 to 8 hours to complete.
2) Intel I7 2620 mobile (2757 float/9841 integer), from 3 to 4 hours to complete.

There is some Intel tricks?
Other volunteers have same experience?
ID: 838 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Col323

Send message
Joined: 5 Oct 15
Posts: 17
Credit: 1,335,501
RAC: 0
Message 839 - Posted: 6 Apr 2016, 15:24:50 UTC
Last modified: 6 Apr 2016, 15:25:31 UTC

Not the same for me.

AMD Phenom(tm) II X6 1065T Processor [Family 16 Model 10 Stepping 0]

Float 2374.07 million ops/sec
Integer 8699.13 million ops/sec

It's done 64 tasks in March & April, all in about 3:45 give or take. This is on Windows 7.
ID: 839 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
[VENETO] boboviz

Send message
Joined: 9 Apr 15
Posts: 171
Credit: 1,371,098
RAC: 1,159
Message 840 - Posted: 6 Apr 2016, 18:14:23 UTC - in response to Message 839.  

It's done 64 tasks in March & April, all in about 3:45 give or take. This is on Windows 7.


Both my pc are Win10 64 bit, but i don't think this is the cause...
ID: 840 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Fuzzy Duck

Send message
Joined: 8 Mar 16
Posts: 7
Credit: 209,249
RAC: 0
Message 841 - Posted: 6 Apr 2016, 18:51:54 UTC - in response to Message 840.  

Seems to be significant difference depending on OS.

My iMac running Darwin 15.4.0 with i7-2600 is completing WU's in under 5k seconds (1h 25m). Amazingly, it is 2nd only to a server with 72 cores in the machine rankings.

http://denis.usj.es/denisathome/top_hosts.php
ID: 841 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile sir spuddly buddly
Avatar

Send message
Joined: 30 Aug 15
Posts: 47
Credit: 1,248,591
RAC: 0
Message 842 - Posted: 6 Apr 2016, 19:07:22 UTC
Last modified: 6 Apr 2016, 19:11:08 UTC

Linux could be the difference and if Hyper-threading slows all the WUs down. Poems WUs were (are still?) a lot quicker with Linux.
update - looking at an old laptop with Fedora it's running a lot quicker then I would guess.
LZ Loon
The Most Handsome Man on the Interweb (TM)
ID: 842 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
[VENETO] boboviz

Send message
Joined: 9 Apr 15
Posts: 171
Credit: 1,371,098
RAC: 1,159
Message 843 - Posted: 7 Apr 2016, 8:06:04 UTC

The next beta batch i'll try a linux virtual machine....
ID: 843 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Greger

Send message
Joined: 10 Apr 15
Posts: 17
Credit: 13,453,681
RAC: 0
Message 844 - Posted: 7 Apr 2016, 16:11:57 UTC - in response to Message 841.  
Last modified: 7 Apr 2016, 16:13:52 UTC

I did run that host with Server 2016 Preview 4 and got some problem with NUMA. It made local server in 2 NUMA groups so i could only use 1 socket.
So 1 core running 2 task at same time result:

Measured floating point speed 3044.79 million ops/sec
Measured integer speed 13763.75 million ops/sec

Latest Beta task
19,268.66 sec (this task was 2 task at first progress ~60% and 1 task last 40 %)
23,545.96 sec (this task was 2 task at first progress ~80% and 1 task last 20 %)

I put in an old SSD in to it with kubuntu from my other server host with Opterons.
So with linux with same hardware, cooler and clockspeed, but old SSD and kubuntu i got
Measured floating point speed 3256.84 million ops/sec
Measured integer speed 11875.85 million ops/sec
Latest Beta task ~7,300-7,400 sec

If i look at same cpu that have 1 socket and same ram memory but Win 10
Measured floating point speed 3000.93 million ops/sec
Measured integer speed 11569.89 million ops/sec
Latest Beta task 14,000-14,300 sec

My old opteron server with Ubuntu 15.10 got
Measured floating point speed 1938.62 million ops/sec
Measured integer speed 7666.42 million ops/sec
Latest Beta task ~14,900-15,500 sec

Summary from my result i can see double performance when i throw in an old SSD with kubuntu to it even if Measured floating point and integer didnĀ“t change that much from win server 2016 and kubuntu.

An Xeon 2696v3 (Clockspeed 2.3 GHz Turbo Speed: 3.6 GHz) will do a same work as an opteron 6276 with linux (Clockspeed: 2.3 GHz, Turbo Speed: 2.6 GHz.
Here is some difference with motheboard, cooling and ram but same typ of SSD.
ID: 844 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
[VENETO] boboviz

Send message
Joined: 9 Apr 15
Posts: 171
Credit: 1,371,098
RAC: 1,159
Message 845 - Posted: 7 Apr 2016, 16:33:08 UTC - in response to Message 844.  

An Xeon 2696v3 (Clockspeed 2.3 GHz Turbo Speed: 3.6 GHz) will do a same work as an opteron 6276 with linux (Clockspeed: 2.3 GHz, Turbo Speed: 2.6 GHz.


If this test is ok, an 5-years-old Opteron crunch as a recent Xeon only with OS change

It seems strange with a non-optimized app
ID: 845 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Greger

Send message
Joined: 10 Apr 15
Posts: 17
Credit: 13,453,681
RAC: 0
Message 846 - Posted: 7 Apr 2016, 16:55:34 UTC - in response to Message 845.  

The SSD have not been used for some time so it use an old version of boinc manager but no optimized app on it now.

The compare as in "runtime per core", there is around 1000 sec if i compare those task. if we look at amount of core there is a different.

Opteron 6276 at 16 cores and Xeon 2696 with HT 36 cores. Small change with task running with hyperthreads.
ID: 846 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile sir spuddly buddly
Avatar

Send message
Joined: 30 Aug 15
Posts: 47
Credit: 1,248,591
RAC: 0
Message 848 - Posted: 8 Apr 2016, 4:49:24 UTC - in response to Message 845.  

An Xeon 2696v3 (Clockspeed 2.3 GHz Turbo Speed: 3.6 GHz) will do a same work as an opteron 6276 with linux (Clockspeed: 2.3 GHz, Turbo Speed: 2.6 GHz.


If this test is ok, an 5-years-old Opteron crunch as a recent Xeon only with OS change

It seems strange with a non-optimized app


As I wrote before some DC projects are quicker with Linux then Windows. It's not the first time I've seen it.
LZ Loon
The Most Handsome Man on the Interweb (TM)
ID: 848 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
[VENETO] boboviz

Send message
Joined: 9 Apr 15
Posts: 171
Credit: 1,371,098
RAC: 1,159
Message 849 - Posted: 8 Apr 2016, 12:07:38 UTC - in response to Message 848.  
Last modified: 8 Apr 2016, 12:10:00 UTC

As I wrote before some DC projects are quicker with Linux then Windows. It's not the first time I've seen it.


Yeap, i know, also rosetta linux app is better than Windows app. But a little difference is ok. Here seem that is a big difference.
Only admins can answer to our assumptions
ID: 849 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile jcastro
Avatar

Send message
Joined: 16 Mar 15
Posts: 219
Credit: 14,859
RAC: 0
Message 850 - Posted: 8 Apr 2016, 12:48:27 UTC
Last modified: 8 Apr 2016, 12:48:36 UTC

Hi!

To compiled beta version of the app we didn't use processor specific compiler. The MAC OS users could have a faster version because we use proprietary compiler from apple that could be optimize.

The Windows version is compiled on a Windows XP virtual machine using Mingw gcc compiler. Linux version is compiled on old scientific linux virtual machine downloaded from official BOINC repository to permit to run in old machines.

The code and compile options are the same to all platforms so the differences could be just on the different compiler used.

Also beta version of the app stores a lot of information in log files, it could have impact on the speed of the tasks ( write to disk speed).

Best regards, Joel.
ID: 850 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile sir spuddly buddly
Avatar

Send message
Joined: 30 Aug 15
Posts: 47
Credit: 1,248,591
RAC: 0
Message 851 - Posted: 9 Apr 2016, 9:15:28 UTC

" The MAC OS users could have a faster version because we use proprietary compiler from apple that could be optimize."
That might explain Darwin linx being faster.
LZ Loon
The Most Handsome Man on the Interweb (TM)
ID: 851 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote

Message boards : Number crunching : Benchmarks of beta