Spectre/Meltdown impact

FAQ, getting help, user experience about PrimoCache
Post Reply
User avatar
Jaga
Contributor
Contributor
Posts: 692
Joined: Sat Jan 25, 2014 1:11 am

Spectre/Meltdown impact

Post by Jaga »

I recently upgraded my primary OS drive to a NVMe, and did some benchmarking to see where it stood with/without Primocache. While doing that, I also decided I'd see where the impact of the Spectre/Meltdown patches stood as far as drive and memory impact for Primocache. The results were interesting.

For reference, my system specs:
- i7-6700k @ 4.8ghz (OC'd)
- 64GB of DDR4 RAM @ 3000mhz (OC'd)
- MSI Z170A Gaming M5 Motherboard
- Samsung 970 EVO 1TB NVMe
- NVidia 2080ti + NVidia 1060ti
- 1000 Watt PSU
- Windows 10 Pro, build 1803 (17134.345)
- Benchmark Software: CrystalDiskMark 5.0.2 x64
- Custom watercooling @ 9C (no load)

I used CrystalDiskMark since the results are easy to interpret, and I used a slightly older version so I could do comparisons with my old SSD benchmarks. And I made sure to launch it with Administrator rights, as not doing so can affect results. I used inSpectre to disable/enable Spectre and Meltdown patches between reboots. Primocache was set with a 32GB L1 cache (4kb blocks aligned with the NVMe volume), with 300s deferred writes.

What I saw was rather surprising, especially considering this post and the assumption that Spectre+Meltdown supposedly don't impact Primocache much with Windows 10. I saw noticeable slowdowns for certain read/write tests when Spectre+Meltdown were enabled via the OS.

With Spectre/Meltdown disabled:

Image


With Spectre/Meltdown enabled:

Image

Note that for most tests, there was an appreciable loss in performance. The only test that didn't suffer was full sequential reads or writes.

Having seen this firsthand now, I'd be curious to hear if Romex Support has done any testing of their own. And to hear if they've had a chance to test the Beta build for the upcoming 2019 W10 patch that supposedly brings Spectre/Meltdown back to near-normal performance. From what I can tell, enabling the Spectre+Meltdown patches in W10 definitely impacts even kernel mode processes, like Primocache.
cichy45
Level 4
Level 4
Posts: 38
Joined: Sun Oct 14, 2018 3:34 pm

Re: Spectre/Meltdown impact

Post by cichy45 »

What was your setup? Did you only disabled OS patches, or you tested on pre-spectre UEFI and after patching?
User avatar
Jaga
Contributor
Contributor
Posts: 692
Joined: Sat Jan 25, 2014 1:11 am

Re: Spectre/Meltdown impact

Post by Jaga »

The only change between the tests was disabling both patches in the OS with inSpectre, and rebooting (a requirement). Both of the above tests were run within ~10 minutes of each other. I do have pre-spectre/meltdown test results, but I've tweaked the system a little since then (higher overclocks now) which would skew a comparison.
jussssx1
Level 3
Level 3
Posts: 18
Joined: Sun Dec 17, 2017 7:19 pm

Re: Spectre/Meltdown impact

Post by jussssx1 »

Yes, the results you got are similar (or even slightly better) than what we have observed. Spectre/Meltdown impacts
latency and that's why fast transactions become significantly slower (e.g. high speed storage, such as primocache or
optane; and fast network io)

-Jussi
User avatar
Jaga
Contributor
Contributor
Posts: 692
Joined: Sat Jan 25, 2014 1:11 am

Re: Spectre/Meltdown impact

Post by Jaga »

Thanks for the confirmation Jussi. I suspected that was the case when a sequential operation didn't suffer any performance loss with the patches enabled, but any random operation did.
User avatar
Support
Support Team
Support Team
Posts: 3623
Joined: Sun Dec 21, 2008 2:42 am

Re: Spectre/Meltdown impact

Post by Support »

PrimoCache works in the kernel mode, so it's native performance shall not be affected. However, benchmark software or other applications run and send I/O requests in the user mode. A simplified I/O processing path is likes below.
1) User applications generates I/O requests, and send to PrimoCache,
2) PrimoCache processes and completes I/O requests,
3) User applications receive I/O results and do additional processing as per applications functions.
Step 1) and 3) run in the user mode, so they may be affected by Spectre/Meltdown patches.
I think that's why you see a performance down with benchmark software.
User avatar
Jaga
Contributor
Contributor
Posts: 692
Joined: Sat Jan 25, 2014 1:11 am

Re: Spectre/Meltdown impact

Post by Jaga »

That makes sense, bottle-necked at a higher level in the OS. There's not much of a fix except waiting for Microsoft to get around to patching the OS (disabling Spectre/Meltdown for daily use isn't recommended). Fortunately there is something in the works due out in 2019.

When I notice the patch arrive I'll re-do the tests for comparison. Thanks for the input Support.
cichy45
Level 4
Level 4
Posts: 38
Joined: Sun Oct 14, 2018 3:34 pm

Re: Spectre/Meltdown impact

Post by cichy45 »

I run my system with disabled meltdown/spectre patches. Norman user should not be scared about someone exploiting them. Maybe if you are running some rack servers in homelab then it is important to have them enabled.
User avatar
Jaga
Contributor
Contributor
Posts: 692
Joined: Sat Jan 25, 2014 1:11 am

Re: Spectre/Meltdown impact

Post by Jaga »

Well, supposedly there are vulnerabilities you can run into even from web pages that exploit one or the other. From that perspective, I'd think servers are even more secure, since browsing/downloading/etc is kept at an absolute minimum on a server. Workstations where web interaction is far higher (including email) would sound like they were more vulnerable.
cichy45
Level 4
Level 4
Posts: 38
Joined: Sun Oct 14, 2018 3:34 pm

Re: Spectre/Meltdown impact

Post by cichy45 »

O, but firstly you have to fall into this trap. I guess if someone is technical enough to use primocache in their setup then he/she is also educated enough not to fall in some phishing or running random pieces execs on system. At least that is my opinion, I know that with my behavior I will be safe without those patches.
If someone work includes more risky behavior like dealing with lots of random emails then it might be easier to sneak up on them so better to have patches enabled.
Post Reply