Hi.
We are looking into move our FPGA build environment so that we can take advantage of the many features that Virtualization has, achieving, golden templating, fast provisioning and mobility between hosts.
The biggest obstacle we face is that Quatus is cpu hungry and will latch onto a number of cores (dependant on the number of process running) leaving the remaining cores mostly utilization.
Hardware spec of host :
Manufacturer: Supermicro
Model: X8DTU
CPUs: 2 x Intel Xeon X5690 @ 3.47 GHz
Memory: 98 GB Ram
HDD: 1TB 7200RM WD1003FBYX
The base comparison for the benchmarking running the build environment natively on the above hardware.
We then ran a single guest machine with 6 vCPS and 80 GB Ram , with no resource pools .
The outcome of this was 2 process build was 20 min longer (typical build is 8 hours 10min +/- 10 min ) compared to native tin and a 4 build processes was 2 hours longer (typical build is 10 hours 10min +/- 10 min ) than on native tin. These results where duplicated in a second set of tests.
we are now testing a single VM in a resources pool with
CPU:
Shares High
Reservation 25000 MHz
Limit 30000MHz
Memory
Shares High
Reservation 87000MB
Limit 88443
The reservation are higher that the expected utilization. Here I'm trying to see what impact the scheduler has on the Guest machine and if there can be any performance gains here.
Based on testing thus far we would look as running 2 guest machines and running 2 builds each thus saving 2 hours when compared to running on native tin.
I would like to be able to do better.
What should I be looking at to get more out of the system ?
What affect would cpu affinity have ?
What would happen If I removed Hyper-threading (Quartus typical doesn't like it) , would the hyper-visor provide faster access to CPU cycles ?
Looking for any suggestions that will help .
Note the majority of the process is in ram during the build process so disk / network I/O are not relevant.
Thank you .