7800X3D lacks SSE
26 Jun 2023 10:45:01 UTC
Topic 229719
(moderation:
Hi everybody, here is something odd:
On my new build in the boinc startup log I can find
"Mon 26 Jun 2023 11:13:14 AM CEST Processor features: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good amd_lbr_v2 nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 x2apic ..."
while the serverlog related to a request for FGRP#5-Tasks contains
"2023-06-26 09:25:45.1985 [PID=306457] [version] Checking plan class 'FGRPSSE'
2023-06-26 09:25:45.2000 [PID=306457] [version] reading plan classes from file '/BOINC/projects/EinsteinAtHome/plan_class_spec.xml'
2023-06-26 09:25:45.2000 [PID=306457] [version] CPU [' family 25 model 97 stepping 2 '] lacks feature ' sse '
2023-06-26 09:25:45.2000 [PID=306457] [version] no app version available: APP#46 (hsgamma_FGRP5) PLATFORM#7 (x86_64-pc-linux-gnu) min_version 0
2023-06-26 09:25:45.2051 [PID=306457] [debug] [HOST#13146320] MSG(high) No work sent"
and accordingly no work is sent. So, BOINC on startup dtermines the CPU has the 'SSE' feature while somewhere in the einstein server request process it is determined the CPU has not. Doesn't seem right to me.
The sytem is is a Debian 12 install right from the repos, no tinkering, LXqt desktop, root/very-unprivileged-user(me) setup using the old (non-graphical) installer.
Interestingly, running the same hardware under Win10 the system received FGRP#5-Work, albeit processing it terribly slow, at least in comparison to my AMD5800X3D-System. Shame on me, at the time I did not record a server log.
So, did anybody experience anything similar or can shed some light on this?
Regards, Walton
it doesn't actually lack the
)
it doesn't actually lack the feature, but the feature list gets truncated for being too long, and the project only sees that truncated list.
one workaround could probably be to take the stock application binary, and form your own app_info.xml file to run it via anonymous platform.
_________________________________________________________________________
Upgrade to a newer BOINC
)
Upgrade to a newer BOINC client. The newer clients don't truncate the cpu feature list.
Upgrade to a newer BOINC
)
Upgrade to a newer BOINC client. The newer clients don't truncate the cpu feature list.
Ryzen 7950x lacks feature sse2
But also requires Einstein to update their server code.
So running the app as an anonymous platform is the workaround until Einstein fixes their server code.
Keith Myers wrote: Upgrade
)
reading down in the comments and notes, this looks like it was a change to the server side code and it wont be fixed by just a client upgrade. DA specifically calls that out.
_________________________________________________________________________
Thank you all, that was
)
Thank you all, that was informative. Taking this in now.
Still, I wonder how that relates to what I observed under W10, as I described it below. I was using BOINC 7.20.2 at the time, for no particular reason other than that it was in my has-worked-well-so-far-toolbox. Given the information you pointed me to, it should have had the same problem, still the Windows-App downloaded and chewed through its workload until I stopped it. So, what's different there?
Cheers,
Walton
Maybe I should clarify one
)
Maybe I should clarify one more thing, just to add to the Windows comparison. Debian 12 installs BOINC 7.20.5 "Pre Release" from the repositories. The BOINC project advertises a different version for Linux x64.
Yet one more thing. "how that
)
Yet one more thing. "how that relates" gives the wrong tone here. What I meant was, what all that means regarding what I observed under Windows 10.
Yes, the quandary here is why
)
Yes, the quandary here is why the difference between OS' when the issue lies primarily with the server code which is agnostic to OS type. I would have expected the same problem under Windows.
But after reviewing the code for the two files that are used to find the cpu feature set, I see they are very different between Windows and Linux.
hostinfo_linux.cpp
hostinfo_win.cpp
There is 10X more code in the Windows file compared to the Linux file.
188 lines for Linux
1692 lines for Windows
I'd still give the newer client releases a shot. May work, you never know.
Any Master release past March 3 when the issue #5123 fix was merged into Master branch.
Thanks Keith. Just to let
)
Thanks Keith.
Just to let you know, I'll stop here though. The systems purpose is to find it's limitations doing everything stock, and I am OK with it just now. However, I think the find actually is relevant to the public, as more users will encounter this, and can find the information here, as well as it is to the project, as ignoring too much CPU power is not good for them, with more AMD 7XXXs coming into play. But that's up to them to decide.
Just now, I will do what I always do, which is load each system with what it can do best, and not load it with what it can't do. If I find time I want to try stock app/app info, and if I do I will not forget to post.
What actually intrigues me is what happens under Windows (Linux seems fairly clear here). If I may try a wild guess: Assuming BOINC suffers the same phenomenon under Windows, can the Einstein-Windows-App default to straightfoward float arithmetics on the CPU to do it's math, if no SSE is determined, so that it would be downloaded and put to work? That would explain long runtimes, and at the same time shadow the string length problem. But to be clear, I am totally making this up, it's not based on any code reading.
Cheers,
Walton
Unlikely for the task to do
)
Unlikely for the task to do that, fallback that is. Unless that path was already coded into the application.
The Einstein devs just need to rewrite the application and update their server software.
I run a 7950X here so would have been bit by this issues too if I ran cpu work. Only running gpu work currently but would have had no issue crafting an anonymous application platform if needed.
I run mostly anonymous anyway for the optimized gpu apps on most projects.