3 ms·

I'm not defending Intel or even particularly have a positive opinion of Intel overall but I don't know how accurate this test is considering the difference in L

I'm not defending Intel or even particularly have a positive opinion of Intel overall but I don't know how accurate this test is considering the difference in Linux kernel versions used in this benchmark. According to the article the Intel system is running a 3.10 kernel (doesn't say which distro but I'm pretty sure it's RHEL 7 or CentOS 7) while the EPYC system is running Ubuntu 19.04 with the latest 5.0 kernel. Also I'm not sure if the 3.10 kernel they're running on the Xeon system has been patched for Meltdown and Spectre which would could also affect the performance. I doubt it would amount to a significant difference in performance but usually benchmarks tests like this use the same OS and software stack for the tests.

rbanffy 7y ago
The performance per dollar is about 5 times better with the AMD box. I'm very sure no new kernel would be able to compensate for that.
- weare138 7y ago
  Hence the reason I said I doubt it would make much of a difference but the 5.0 kernel generally outperforms 3.10. 3.10 has has already been deprecated and isn't updated anymore. The EPYC 7742's are still going to trounce the Xeon's but the Xeon's would probably have better real world performance on the latest 5.0 kernel and a recent distro versus kernel 3.10 and a dated distro like RHEL/CentOS 7. As I mentioned before, typically the same version of OS is used in these types of benchmarks.
  - bubblethink 7y ago
    RHEL backports the hell out of their kernels. They are monstrosities. The would not make any money if their OS performed significantly worse than the competition. Remember how much RHEL costs. And phoronix routinely does these performance tests across kernel versions. There is not that much difference. I think in one of his recent tests he even went back to ubuntu 12.04.
simcop2387 7y ago
This might actually be a win for intel, 3.10 may not have all the spectre and meltdown patches backported to it (though I suspect if it's RHEL 7 that it does).
pizza234 7y ago
Valid point, although a few considerations: 1. generally speaking, kernel version increases don't translate to performance improvements (example: https://www.phoronix.com/scan.php?page=article&item=linux-50-sliding&num=1 https://www.phoronix.com/scan.php?page=article&item=linux-50...) 2. the Dell 840 are latest generation servers (https://www.dell.com/en-us/work/shop/poweredge-rack-servers/sf/poweredge-rack-servers https://www.dell.com/en-us/work/shop/poweredge-rack-servers/...), so I doubt that they're not optimized to be competitive with latest operating systems). after all, this is a real-world machine one buys, so it's supposed to be highly performing. All in all, the performance improvement seems to be proportional to the single thread performance improvement (however small) componded by the higher core count. However of course, it can't be excluded that the difference is at least partially caused by the different setup.
- weare138 7y ago
  The benchmark you linked to somewhat misses the point I'm trying to make. It's specifically comparing the 4.0 branch (4.12-4.20) to 5.0 on the same distribution. The distribution is also a factor in performence. Check out this benchmark comparing CentOS 7 to other distributions on the same hardware: https://www.phoronix.com/scan.php?page=article&item=5distros-post-spectre&num=2 https://www.phoronix.com/scan.php?page=article&item=5distros... CentOS 7 is slower in most of the tests compared to a recent version of Ubuntu (17.10 was used in this comparison), which itself is slower than Clear Linux in most tests.
  - martinpw 7y ago
    I had no idea the performance differences between the distros was so significant. Which settings/compiler flags have the largest effect here?

rbanffy 7y ago

The performance per dollar is about 5 times better with the AMD box. I'm very sure no new kernel would be able to compensate for that.

weare138 7y ago
Hence the reason I said I doubt it would make much of a difference but the 5.0 kernel generally outperforms 3.10. 3.10 has has already been deprecated and isn't updated anymore. The EPYC 7742's are still going to trounce the Xeon's but the Xeon's would probably have better real world performance on the latest 5.0 kernel and a recent distro versus kernel 3.10 and a dated distro like RHEL/CentOS 7. As I mentioned before, typically the same version of OS is used in these types of benchmarks.
- bubblethink 7y ago
  RHEL backports the hell out of their kernels. They are monstrosities. The would not make any money if their OS performed significantly worse than the competition. Remember how much RHEL costs. And phoronix routinely does these performance tests across kernel versions. There is not that much difference. I think in one of his recent tests he even went back to ubuntu 12.04.

simcop2387 7y ago

This might actually be a win for intel, 3.10 may not have all the spectre and meltdown patches backported to it (though I suspect if it's RHEL 7 that it does).

pizza234 7y ago

Valid point, although a few considerations: 1. generally speaking, kernel version increases don't translate to performance improvements (example: https://www.phoronix.com/scan.php?page=article&item=linux-50-sliding&num=1 https://www.phoronix.com/scan.php?page=article&item=linux-50...) 2. the Dell 840 are latest generation servers (https://www.dell.com/en-us/work/shop/poweredge-rack-servers/sf/poweredge-rack-servers https://www.dell.com/en-us/work/shop/poweredge-rack-servers/...), so I doubt that they're not optimized to be competitive with latest operating systems). after all, this is a real-world machine one buys, so it's supposed to be highly performing. All in all, the performance improvement seems to be proportional to the single thread performance improvement (however small) componded by the higher core count. However of course, it can't be excluded that the difference is at least partially caused by the different setup.

weare138 7y ago
The benchmark you linked to somewhat misses the point I'm trying to make. It's specifically comparing the 4.0 branch (4.12-4.20) to 5.0 on the same distribution. The distribution is also a factor in performence. Check out this benchmark comparing CentOS 7 to other distributions on the same hardware: https://www.phoronix.com/scan.php?page=article&item=5distros-post-spectre&num=2 https://www.phoronix.com/scan.php?page=article&item=5distros... CentOS 7 is slower in most of the tests compared to a recent version of Ubuntu (17.10 was used in this comparison), which itself is slower than Clear Linux in most tests.
- martinpw 7y ago
  I had no idea the performance differences between the distros was so significant. Which settings/compiler flags have the largest effect here?