Dienstag, 20. März 2018

real virtual CPUs

Some software changes it's behavior based on capabilities of the system it's running on.
but sometimes it's interesting to check how a software would heave on a different system, which is not at hand right now.

On Linux, a lot of information about the current system can be found in /proc and /sys.
These filesystems are virtual, so they can not changed easily with an editor.

In my case I want to simulate  a lot more CPUs.
These are visible in several locations.
The most know is probably /proc/cpuinfo.  There you find a block of information for each CPU the kernel knows about. Based on the current configuration, I create a modified fake file somewhere in a different space:
# cpus.sh


for ((soc=0;soc<max_socket;soc++)); do
    for (( cor=0;cor<max_core;cor++)); do
echo "processor       : $count
vendor_id       : GenuineIntel
cpu family      : 6
model           : 37
model name      : Intel(R) Xeon(R) CPU E5-2660 v3 @ 2.60GHz
stepping        : 1
microcode       : 0x3b
cpu MHz         : 2596.103
cache size      : 25600 KB
physical id     : $soc
siblings        : $max_core
core id         : $cor
cpu cores       : $max_core
apicid          : 0
initial apicid  : 0
fpu             : yes
fpu_exception   : yes
cpuid level     : 11
wp              : yes
flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts mmx fxsr sse sse2 ss syscall nx rdtscp lm constant_tsc arch_perfmon pebs bts nopl xtopology tsc_reliable nonstop_tsc aperfmperf pni pclmulqdq ssse3 cx16 sse4_1 sse4_2 x2apic popcnt tsc_deadline_timer aes hypervisor lahf_lm ida arat epb pln pts dtherm pti tsc_adjust
bugs            : cpu_meltdown spectre_v2
bogomips        : 5193.98
clflush size    : 64
cache_alignment : 64
address sizes   : 42 bits physical, 48 bits virtual
power management:

    let count=count+1
and create a file with ./cpus.sh>cpuinfo.256

There is another location as well: /sys/devices/system/cpu.
In this directory are several directories and files with interesting information.

I copy the fill directory to another place (ignoring all the errors).
First the number of cpu[id] directories might need adjustment.
In my case a simple set of symlinks is sufficient:
for i in {2..255} ; do
  echo $i
  ln -s cpu1 cpu${i}
In every cpu[id] durectory there is a symlinik to which node it belongs: node0 -> ../../node/node0
So it might be required to spoof proper entries in /sys/devices/system/node. In my case it's not required.

The last fix required in my case is in the file cpu/online.
It contains 0-255 now (instead of 0-2).

As I mentioned above the original files can not be manipulated as they are not real files.
The mount option --bind does the trick:
mount --bind <my_working_dir>/cpuinfo.256 /proc/cpuinfo
mount --bind <my_working_dir>/cpu /sys/devices/system/cpu

After these nice manipulations, my sandbox Oracle instance shows now plenty of CPUs:
SQL> show parameter cpu_count                                                                                                                                                      

NAME                                 TYPE        VALUE
------------------------------------ ----------- ------------------------------
cpu_count                            integer     256

Update (2018-03-21):
For Oracle Databases I got 2 hints how to make it calculate with more CPUs than really available.

with this small stap script:
function modify_rax() %{ long ret; ret = 6; memcpy( ((char *)CONTEXT->uregs) + 80 , &ret, sizeof(ret)); %}
probe process(“oracle”).function(“skgpnumcpu”).return { modify_rax(); }


Kommentar veröffentlichen