I got it, it's in the the libgcc64-32bit: libgcc_s.so.1
Can this be used?
Probably. The new 1.15 App is meant to shed a bit more light on this, it tries to load the library manually and logs the error if it can't.
Quote:
I repeat:
Windows with 32bit App: 12,001.28 CPU time: 11,861.91
Linux X64 on the same host: 13,123.66 Cpu time: 13,052.07
It's i7 .
Is that nothing? On other hosts it might be remarkable slower.
Well, I did run the very same workunit in parallel on two cores of the same (Linux) machine, standalone (i.e. without BOINC communication) and with the machine being otherwise empty. Did you? With your comparison there might be differences in the workunit, the I/O when reading the data, system's power saving settings etc. That doesn't mean your comparison is wrong, I just wouldn't trust it as much as mine.
Anyway, I'll see what the 1.15 App tells me and then re-configure the apps to the old behavior.
I got a dual boot system. Win 7 as game loader and for tax(Elster).
One time with Win 7 64bit and the other time on Linux 64bit, on the same host.
No power saving, just full speed, at max. performance.
I did some Math on the same System:
Linux(38 results): 13244,0047368421 / 13164,5931578947
Windows(20 results): 11242,656 / 11219,681
The last Value is the CPU-time and I must say that I play "The Elder Scrolls V: Skyrim" on Win 7 sometimes, so the values might be worse.
In the 1.16 Linux App I implemented a workaround for the "pthread_cancel" problem.
I changed the configuration of the X64 plan class, such that clients that report that they can run 32Bit Apps (i.e. that report "i686-pc-linux-gnu" as alternative platform) should get the 1.16 32Bit Linux App.
I'll take a look at the WU runtimes of such clients in a week.
FWIW there are currently 2388 computers that returned results from both apps (1.13 X64 and 1.16 SSE2). The difference in run time averaged over these 2388 computers is 120s (CPU time) / 43.5s (elapsed time). With average runtimes of 23085s (CPU) / 24508s (elapsed) the difference is <0.5%.
FWIW there are currently 2388 computers that returned results from both apps (1.13 X64 and 1.16 SSE2). The difference in run time averaged over these 2388 computers is 120s (CPU time) / 43.5s (elapsed time). With average runtimes of 23085s (CPU) / 24508s (elapsed) the difference is <0.5%.
BM
This is not true:
My i7 2600K(4,2GHz) with the 64bit App used:13244,0047368421 / 13164,5931578947
My i7 2600K(4,2GHz) with the 32bit App uses: 16845,9825 / 16373,2875
The 64bit App is far faster!
This may be different on your particular computer for various reasons related to OS and its settings, other running processes or whatever.
My numbers are taken from the E@H database, averaged over >2000 computers and roughly 50000 tasks, and these are consistent with the result of a single previous experiment under tightly controlled conditions.
In what sense would you call these numbers "not true"? Do you think I'm making these up? What purpose would that serve?
I think I will switch to Windows 7, if you you don't respect these facts.
Sorry for calling you a lier.
Bernd, the 64bit app is far faster, see my numbers! That is true and BRPS tasks take in Win 7 about 25 minutes.
I got a duel boot system.
Did you look in the thead, all values are from Linux.
The 6-core AMD had a cooling problem, but the first 32bit WU took around 7 hours. A long time. I still have an XP license, that I can install on my 6-core AMD.
I said I wasn't at home the whole day. No other processes, I have even quit T4T(with very limited CPU resources) on my Linux host. 30 minutes for a BRPS task, that's far to long, compared to 24 minutes.
I didn't compare the times on my root server jet.
A new record, while I write this 4:56:22 for a SSE2 app.
I will keep you informed.
You are certainly free to do what works best for you.
What I can say is that the behavior of the 32 vs 64 Bit Linux Apps is highly atypical on your host; it deviates a lot from the average:
[pre]
mysql> select (avg(x64.cpu)+avg(x86.cpu))/2.0 as avg_cpu, avg(x64.cpu-x86.cpu) as avg_cpu_diff, (avg(x64.el)+avg(x86.el))/2.0 as avg_elapsed, avg(x64.el-x86.el) as avg_el_diff, count(*) as hosts from (select hostid, avg(cpu_time) as cpu, avg(elapsed_time) as el from result where appid=20 and server_state=5 and outcome=1 and app_version_id = 331 group by hostid) as x64 inner join (select hostid, avg(cpu_time) as cpu, avg(elapsed_time) as el from result where appid=20 and server_state=5 and outcome=1 and app_version_id = 350 group by hostid) as x86 on x64.hostid = x86.hostid;
+-----------------+-----------------+-----------------+-----------------+-------+
| avg_cpu | avg_cpu_diff | avg_elapsed | avg_el_diff | hosts |
+-----------------+-----------------+-----------------+-----------------+-------+
| 23257.374359047 | 105.74232885702 | 24707.581969798 | 71.425287120414 | 2477 |
+-----------------+-----------------+-----------------+-----------------+-------+
mysql> select (avg(x64.cpu)+avg(x86.cpu))/2.0 as avg_cpu, avg(x64.cpu-x86.cpu) as avg_cpu_diff, (avg(x64.el)+avg(x86.el))/2.0 as avg_elapsed, avg(x64.el-x86.el) as avg_el_diff, count(*) as hosts from (select hostid, avg(cpu_time) as cpu, avg(elapsed_time) as el from result where appid=20 and server_state=5 and outcome=1 and app_version_id = 331 group by hostid) as x64 inner join (select hostid, avg(cpu_time) as cpu, avg(elapsed_time) as el from result where appid=20 and server_state=5 and outcome=1 and app_version_id = 350 group by hostid) as x86 on x64.hostid = x86.hostid where x64.hostid = 737731;
+-----------------+------------------+-----------------+------------------+-------+
| avg_cpu | avg_cpu_diff | avg_elapsed | avg_el_diff | hosts |
+-----------------+------------------+-----------------+------------------+-------+
| 14924.120282051 | -3594.7154358974 | 15394.036525331 | -4379.4286324051 | 1 |
+-----------------+------------------+-----------------+------------------+-------+
[/pre]
What I can't tell you is why.
But to judge the performance of an application it's certainly better to watch its behavior on ~2500 different computers and average over that than to watch its behavior on a single computer only, and under varying conditions (input data for one, probably room temperature for another).
Again, you are certainly free to do what you found working best for your machine; the same way that I will do what works best for the average machine, thus for the average participant and for the project.
RE: I got it, it's in the
)
Probably. The new 1.15 App is meant to shed a bit more light on this, it tries to load the library manually and logs the error if it can't.
Well, I did run the very same workunit in parallel on two cores of the same (Linux) machine, standalone (i.e. without BOINC communication) and with the machine being otherwise empty. Did you? With your comparison there might be differences in the workunit, the I/O when reading the data, system's power saving settings etc. That doesn't mean your comparison is wrong, I just wouldn't trust it as much as mine.
Anyway, I'll see what the 1.15 App tells me and then re-configure the apps to the old behavior.
BM
BM
I got a dual boot system. Win
)
I got a dual boot system. Win 7 as game loader and for tax(Elster).
One time with Win 7 64bit and the other time on Linux 64bit, on the same host.
No power saving, just full speed, at max. performance.
I did some Math on the same System:
Linux(38 results): 13244,0047368421 / 13164,5931578947
Windows(20 results): 11242,656 / 11219,681
The last Value is the CPU-time and I must say that I play "The Elder Scrolls V: Skyrim" on Win 7 sometimes, so the values might be worse.
[edit]
One host Linux: http://einsteinathome.org/host/737731/tasks&offset=140&show_names=0&state=0
One host Windows: http://einsteinathome.org/host/3960976/tasks
In the 1.16 Linux App I
)
In the 1.16 Linux App I implemented a workaround for the "pthread_cancel" problem.
I changed the configuration of the X64 plan class, such that clients that report that they can run 32Bit Apps (i.e. that report "i686-pc-linux-gnu" as alternative platform) should get the 1.16 32Bit Linux App.
I'll take a look at the WU runtimes of such clients in a week.
BM
BM
I got one one on my 6-core
)
I got one one on my 6-core AMD!
ON all other hosts niente.
Hope it performs well.
The interesting result will be on my i7 2600K!
Bernd thanks allot!
FWIW there are currently 2388
)
FWIW there are currently 2388 computers that returned results from both apps (1.13 X64 and 1.16 SSE2). The difference in run time averaged over these 2388 computers is 120s (CPU time) / 43.5s (elapsed time). With average runtimes of 23085s (CPU) / 24508s (elapsed) the difference is <0.5%.
BM
BM
RE: FWIW there are
)
This is not true:
My i7 2600K(4,2GHz) with the 64bit App used:13244,0047368421 / 13164,5931578947
My i7 2600K(4,2GHz) with the 32bit App uses: 16845,9825 / 16373,2875
The 64bit App is far faster!
Look in this Thread(German):
http://www.heise.de/foren/S-Re-Planierraupenfahrer-hier-und-CUDA-Frage-nebenher/forum-74338/msg-21875798/read/
I still have to check the other hosts.
It's even worse: 17482,284 /
)
It's even worse: 17482,284 / 16688,622
Even the BRPS suddenly take 30 minutes where they used to take around 24 minutes: On a GTX 570 oc.
[edit]typos
I wasn't even at home during the day!
RE: This is not
)
This may be different on your particular computer for various reasons related to OS and its settings, other running processes or whatever.
My numbers are taken from the E@H database, averaged over >2000 computers and roughly 50000 tasks, and these are consistent with the result of a single previous experiment under tightly controlled conditions.
In what sense would you call these numbers "not true"? Do you think I'm making these up? What purpose would that serve?
BM
BM
I think I will switch to
)
I think I will switch to Windows 7, if you you don't respect these facts.
Sorry for calling you a lier.
Bernd, the 64bit app is far faster, see my numbers! That is true and BRPS tasks take in Win 7 about 25 minutes.
I got a duel boot system.
Did you look in the thead, all values are from Linux.
The 6-core AMD had a cooling problem, but the first 32bit WU took around 7 hours. A long time. I still have an XP license, that I can install on my 6-core AMD.
I said I wasn't at home the whole day. No other processes, I have even quit T4T(with very limited CPU resources) on my Linux host. 30 minutes for a BRPS task, that's far to long, compared to 24 minutes.
I didn't compare the times on my root server jet.
A new record, while I write this 4:56:22 for a SSE2 app.
I will keep you informed.
You are certainly free to do
)
You are certainly free to do what works best for you.
What I can say is that the behavior of the 32 vs 64 Bit Linux Apps is highly atypical on your host; it deviates a lot from the average:
[pre]
mysql> select (avg(x64.cpu)+avg(x86.cpu))/2.0 as avg_cpu, avg(x64.cpu-x86.cpu) as avg_cpu_diff, (avg(x64.el)+avg(x86.el))/2.0 as avg_elapsed, avg(x64.el-x86.el) as avg_el_diff, count(*) as hosts from (select hostid, avg(cpu_time) as cpu, avg(elapsed_time) as el from result where appid=20 and server_state=5 and outcome=1 and app_version_id = 331 group by hostid) as x64 inner join (select hostid, avg(cpu_time) as cpu, avg(elapsed_time) as el from result where appid=20 and server_state=5 and outcome=1 and app_version_id = 350 group by hostid) as x86 on x64.hostid = x86.hostid;
+-----------------+-----------------+-----------------+-----------------+-------+
| avg_cpu | avg_cpu_diff | avg_elapsed | avg_el_diff | hosts |
+-----------------+-----------------+-----------------+-----------------+-------+
| 23257.374359047 | 105.74232885702 | 24707.581969798 | 71.425287120414 | 2477 |
+-----------------+-----------------+-----------------+-----------------+-------+
mysql> select (avg(x64.cpu)+avg(x86.cpu))/2.0 as avg_cpu, avg(x64.cpu-x86.cpu) as avg_cpu_diff, (avg(x64.el)+avg(x86.el))/2.0 as avg_elapsed, avg(x64.el-x86.el) as avg_el_diff, count(*) as hosts from (select hostid, avg(cpu_time) as cpu, avg(elapsed_time) as el from result where appid=20 and server_state=5 and outcome=1 and app_version_id = 331 group by hostid) as x64 inner join (select hostid, avg(cpu_time) as cpu, avg(elapsed_time) as el from result where appid=20 and server_state=5 and outcome=1 and app_version_id = 350 group by hostid) as x86 on x64.hostid = x86.hostid where x64.hostid = 737731;
+-----------------+------------------+-----------------+------------------+-------+
| avg_cpu | avg_cpu_diff | avg_elapsed | avg_el_diff | hosts |
+-----------------+------------------+-----------------+------------------+-------+
| 14924.120282051 | -3594.7154358974 | 15394.036525331 | -4379.4286324051 | 1 |
+-----------------+------------------+-----------------+------------------+-------+
[/pre]
What I can't tell you is why.
But to judge the performance of an application it's certainly better to watch its behavior on ~2500 different computers and average over that than to watch its behavior on a single computer only, and under varying conditions (input data for one, probably room temperature for another).
Again, you are certainly free to do what you found working best for your machine; the same way that I will do what works best for the average machine, thus for the average participant and for the project.
BM
BM