RFR(XXS): JDK-8007069 VM_Version_x86 cores per cpu is wrong for cpuid 0xb topology case

Markus Grönlund markus.gronlund at oracle.com
Wed Jan 30 13:24:01 PST 2013 

Thanks a lot for pointing me to this Micke!

That is a very interesting ** indeed, "must not use EBX[15:0] to enumerate processor topology of the system"

I guess what is referred to here is exactly what that Vladimir just showed, that the value in EBX[15:0] is not an accurate number of cores on a CPU package. Instead EAX[4:0] should be used to find the level shifts for creating the mask widths for proper sub id extraction (SMT, Core, Pkg) from the 2xAPICIDs when doing topology enumeration.
 
Thanks again Mikael

/Markus




-----Original Message-----
From: Mikael Vidstedt 
Sent: den 30 januari 2013 21:56
To: Vladimir Kozlov
Cc: Markus Grönlund; hotspot-dev at openjdk.java.net
Subject: Re: RFR(XXS): JDK-8007069 VM_Version_x86 cores per cpu is wrong for cpuid 0xb topology case


For completeness, I was looking at the instruction set reference (from January 2013), page 3-152 and 3-153, and specifically the double star
("**") note for the EBX value.

http://download.intel.com/products/processor/manual/253666.pdf

Cheers,
Mikael

On 2013-01-30 11:26, Vladimir Kozlov wrote:
> You are welcome.
>
> Yes, you have to read "small print" not only on financial documents 
> but also in hw architecture manuals :(
>
> Regards,
> Vladimir
>
> On 1/30/13 11:09 AM, Markus Grönlund wrote:
>> Vladimir,
>>
>> many thanks giving me a good example, I have only had 
>> _cpuid_info.tpl_cpuidB0_ebx.bits.logical_cpus = 1, so far.
>>
>> You are right, thank you for showing me this, I did not read the 
>> details on the Note2 on Core level well enough:
>>
>> "2. The number of factory-configured logical processors at this level 
>> is equal to the number of factory-configured cores * the number of 
>> factory-configured logical processors per core."
>>
>> It was my (mis)understanding that Core level did not multiply up to 
>> the SMT level(but id does as the note gives as well as your example 
>> explains).
>>
>> Many thanks Vladimir
>>
>> Cheers
>> Markus
>>
>>
>>
>> -----Original Message-----
>> From: Vladimir Kozlov
>> Sent: den 30 januari 2013 19:58
>> To: Markus Grönlund
>> Cc: Mikael Vidstedt; hotspot-dev at openjdk.java.net
>> Subject: Re: RFR(XXS): JDK-8007069 VM_Version_x86 cores per cpu is 
>> wrong for cpuid 0xb topology case
>>
>> No, the example I showed has HT and VM gives correct results:
>>
>> tpl_cpuidB0_ebx.bits.logical_cpus = 2 
>> tpl_cpuidB1_ebx.bits.logical_cpus = 16
>>
>> (8 cores per cpu, 2 threads per core)
>>
>> Vladimir
>>
>> On 1/30/13 10:47 AM, Markus Grönlund wrote:
>>> Hi Vladimir,
>>>
>>> Yes I have noticed that it will give ok results (so far). For 
>>> example, on my Intel(R) Xeon(R) CPU E5-2690 0 @ 2.90GHz, 
>>> (Sandybridge-EP).
>>>
>>> I have 8 hw threads (from os::processor_count())
>>>
>>> Also, it's using CPUID 0xb (supports topology, 2xAPICID)
>>>
>>> My values are:
>>>
>>> _cpuid_info.tpl_cpuidB0_ebx.bits.logical_cpus = 1 <<-- threads per 
>>> core _cpuid_info.tpl_cpuidB1_ebx.bits.logical_cpus = 8 <<-- cores 
>>> per cpu package
>>>
>>> Current (wrong) logic does:
>>>
>>> No of cores per cpu:
>>>
>>> Returns : _cpuid_info.tpl_cpuidB1_ebx.bits.logical_cpus = 8 / 
>>> _cpuid_info.tpl_cpuidB0_ebx.bits.logical_cpus = 1 == no of cores ==
>>> 8 (correct).
>>>
>>> However,
>>>
>>> This does not account for CPUs that have HT enabled, i.e  if HT is 
>>> enabled, you will get this:
>>>
>>> _cpuid_info.tpl_cpuidB0_ebx.bits.logical_cpus = 2 <<-- threads per 
>>> core _cpuid_info.tpl_cpuidB1_ebx.bits.logical_cpus = 8 <<-- cores 
>>> per cpu package
>>>
>>>
>>> Returns : _cpuid_info.tpl_cpuidB1_ebx.bits.logical_cpus = 8 / 
>>> _cpuid_info.tpl_cpuidB0_ebx.bits.logical_cpus = 2 == no of cores is 
>>> now 4, but the real number of cores is 8.
>>>
>>> So, avoid doing the unnecessary divide operation
>>> (no_of_cores_per_pkg) / (no_of_threads_per_core). The correct info 
>>> is already there. Just return it.
>>>
>>> Return: _cpuid_info.tpl_cpuidB1_ebx.bits.logical_cpus // 8 correct
>>>
>>>
>>> Cheers
>>> Markus
>>>
>>>
>>>
>>> -----Original Message-----
>>> From: Vladimir Kozlov
>>> Sent: den 30 januari 2013 19:33
>>> To: Markus Grönlund
>>> Cc: Mikael Vidstedt; hotspot-dev at openjdk.java.net
>>> Subject: Re: RFR(XXS): JDK-8007069 VM_Version_x86 cores per cpu is 
>>> wrong for cpuid 0xb topology case
>>>
>>> Markus,
>>>
>>> Current code is correct. In what case it is not correct?
>>>
>>> bash-4.1$ psrinfo -pv
>>> The physical processor has 8 cores and 16 virtual processors (0-7
>>> 16-23)
>>>      The core has 2 virtual processors (0 16)
>>>      The core has 2 virtual processors (1 17)
>>>      The core has 2 virtual processors (2 18)
>>>      The core has 2 virtual processors (3 19)
>>>      The core has 2 virtual processors (4 20)
>>>      The core has 2 virtual processors (5 21)
>>>      The core has 2 virtual processors (6 22)
>>>      The core has 2 virtual processors (7 23)
>>>        x86 (GenuineIntel 206D6 family 6 model 45 step 6 clock 2693 MHz)
>>>          Intel(r) Xeon(r) CPU E5-2680 0 @ 2.70GHz
>>>
>>> ash-4.1$ java -XX:+PrintMiscellaneous -XX:+Verbose -version 
>>> [SafePoint Polling address: 0xfffffd7fff1e0000] [Memory Serialize  Page address:
>>> 0xfffffd7fff1d0000] Logical CPUs per core: 2
>>> UseSSE=4  UseAVX=1  UseAES=1
>>> Allocation prefetching: PREFETCHNTA at distance 192, 4 lines of 64 
>>> bytes PrefetchCopyIntervalInBytes 576 PrefetchScanIntervalInBytes 
>>> 576 PrefetchFieldsAhead 1 ContendedPaddingWidth 128 CPU:total 32 (8 
>>> cores per cpu, 2 threads per core) family 6 model 45 stepping 6, 
>>> cmov, cx8, fxsr, mmx, sse, sse2, sse3, ssse3, sse4.1, sse4.2, 
>>> popcnt, avx, aes, ht, tsc, tscinvbit, tscinv
>>>
>>> Thanks,
>>> Vladimir
>>>
>>>
>>> On 1/30/13 9:45 AM, Markus Grönlund wrote:
>>>> Hi Mikael,
>>>>
>>>> I wonder if you are looking at the right section in the manual (and 
>>>> which manual?) here?
>>>>
>>>> I am using information in Intel Application note 485:
>>>>
>>>> http://www.intel.com/content/www/us/en/processors/processor-identif
>>>> ic
>>>> a
>>>> tion-cpuid-instruction-note.html
>>>>
>>>> Section 5.1.12 page 41 describes CPUID 0xb. Essentially, together 
>>>> with EAX=0xb, you pass in subleafs (0, 1, 2...) in ECX for the 
>>>> different levels in the cpu topology you are interested in. Results 
>>>> are given in EBX[15:0] [31:16] == reserved.
>>>>
>>>>
>>>> "There's also a cores_per_cpu in CPUID 04h EAX[31:26] (aka. 
>>>> dcp_cpuid4_eax.bits.cores_per_cpu), do you happen to know what the 
>>>> difference is? Is that maybe a better candidate for this?"
>>>>
>>>> CPUID04 EAX[31:26] is giving you the number of cores [per cpu 
>>>> package] using the legacy "initial APICID (8-bits)" systems. This 
>>>> combined with CPUID 01[logical threads per cpu package] is the 
>>>> legacy way of cpu identification. It is cannot give you as much 
>>>> info as topology however.
>>>>
>>>> Processors supporting CPUID 0xb is using 3-level 2xAPICID (32-bit) 
>>>> for identification (SMT, Core, Pkg), and does not use "initial 
>>>> APICID".
>>>>
>>>> So:
>>>>
>>>> EAX=0xb, ECX=0 == EBX[15:0] == no_of_logical_threads_per_core == 
>>>> (stored in _cpuid_info.tpl_cpuidB0_ebx.bits.logical_cpus)
>>>> EAX=0xb, ECX=1 == EBX[15:0] == no_of_cores_per_pkg == (stored in
>>>> _cpuid_info.tpl_cpuidB1_ebx.bits.logical_cpus)
>>>> ...
>>>>
>>>> Hope this clarifies
>>>>
>>>> Cheers
>>>> Markus
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> -----Original Message-----
>>>> From: Mikael Vidstedt
>>>> Sent: den 30 januari 2013 17:30
>>>> To: Markus Grönlund
>>>> Cc: hotspot-dev at openjdk.java.net
>>>> Subject: Re: RFR(XXS): JDK-8007069 VM_Version_x86 cores per cpu is 
>>>> wrong for cpuid 0xb topology case
>>>>
>>>>
>>>> Looking at the Intel manual for CPUID 0Bh it basically says that 
>>>> one should not use the value in EBX[15:0] for anything other than 
>>>> "...display/diagnostic purposes". There's also a cores_per_cpu in 
>>>> CPUID 04h EAX[31:26] (aka. dcp_cpuid4_eax.bits.cores_per_cpu), do 
>>>> you happen to know what the difference is? Is that maybe a better 
>>>> candidate for this?
>>>>
>>>> Cheers,
>>>> Mikael
>>>>
>>>> On 1/29/2013 2:10 AM, Markus Grönlund wrote:
>>>>> Hi again,
>>>>>
>>>>> Apologies for posting the wrong webrev link, this is the one I 
>>>>> wanted to submit:
>>>>>
>>>>> Bug: https://jbs.oracle.com/bugs/browse/JDK-8007069
>>>>> Webrev: http://cr.openjdk.java.net/~mgronlun/8007069/webrev01/
>>>>>
>>>>>
>>>>> Thanks Mikael for the heads up
>>>>>
>>>>> Cheers
>>>>> Markus
>>>>>
>>>>>
>>>>>
>>>>> -----Original Message-----
>>>>> From: Mikael Gerdin
>>>>> Sent: den 29 januari 2013 10:59
>>>>> To: Markus Grönlund
>>>>> Cc: hotspot-dev at openjdk.java.net
>>>>> Subject: Re: RFR(XXS): JDK-8007069 VM_Version_x86 cores per cpu is 
>>>>> wrong for cpuid 0xb topology case
>>>>>
>>>>> Markus,
>>>>>
>>>>> On 2013-01-29 10:54, Markus Grönlund wrote:
>>>>>> Hi all,
>>>>>>
>>>>>>
>>>>>>
>>>>>> Kindly asking for review for the following very simple change in 
>>>>>> the logic for number of cores for CPUID 0xb topology.
>>>>>>
>>>>>>
>>>>>>
>>>>>> Bug: https://jbs.oracle.com/bugs/browse/JDK-8007069
>>>>>>
>>>>>> Webrev:
>>>>>> http://wikifiles.se.oracle.com/dev/mgronlun/webrevs/8007069/webre
>>>>>> v0
>>>>>> 1
>>>>>> /
>>>>> This URL is not available from outside Oracle.
>>>>>
>>>>> /Mikael
>>>>>
>>>>>>
>>>>>>
>>>>>> Also need someone to sponsor this putback.
>>>>>>
>>>>>>
>>>>>>
>>>>>> Thanks in advance
>>>>>>
>>>>>> Markus
>>>>>>
>>>>

More information about the hotspot-dev mailing list