Question Optane storage performance with & without hyperthreading. Why large differences in some benchmarks but not others?

germanium

Distinguished
Jul 8, 2006
93
9
18,565
I see in some cases huge differences in benchmarks with hyperthreading on compared to with hyperthreading off. I was wondering if anyone can explain this phenominon?

For example Crystal disc mark 4K QD8 thread 8 better than 2300 MB/s without hyperthreading but 1300-1400 MB/s with.

However anvil storage benchmark shows no such huge gain in multithreaded benchmarks & stays around 1700 both with & without hyperthreading on 4K QD16.

These are scores with three Optane Memory drives connected to the PCI-E X16_1 that goes directly to CPU.

Platform is Asus Z390 chipset motherboard with Intel Core I9 9900K CPU.
 
Last edited:

Lutfij

Titan
Moderator
It's customary to include your specs in the body of your thread as opposed to the sig space since that area can and will change over time making this thread and it's solutions moot to anyone in the same boat as you(with the same specs). In listing specs, list them like so:
CPU:
Motherboard:
Ram:
SSD/HDD:
GPU:
PSU:
Chassis:
OS:

If I may ask, are you striving for benchmark results or are you experiencing real world issues with your generally used app's/games?
 

germanium

Distinguished
Jul 8, 2006
93
9
18,565
The specs have not changed since I made the signature specs in the signature & are no more than 2 months old with the exception of updated operating system.

I don't play games on it. This was never meant to be a gaming rig in spite of the motherboard having gaming in the name. Just a highly responsive computer. Nothing I actually do on it stresses it except benchmarks there are no current issues at current clock speeds & voltages with the benchmarks I typically run. The only thing changed was the operating system has been updated to the 1903 may prerelease version of Windows 10. This did negatively effect the storage scores ever so slightly.

I also forgot to add that I am running an Enermax liqmax 240 mm aio liquid cooler. This was true when I made the signature so no change.

My cpu is set to not start down clock until it has exceeded 180 watts for more than 270 seconds & not to exceed 200 watts at all though it does take a moment to react. These are my BIOS settings

All core boost is 5GHz @1.35 volts.

Power supply upgraded to Corsair CX750M to better handle huge currents & operate in its most efficient range at full load. Old power supply was adequate but just barely & made no difference in actual performance so was not listed due to no change in performance.

Hyperthreading is currently turned off but I do activate it when running cpu benchmarks & off otherwise at least until I find a program that benefits from hyperthreading.I already have so many real cores that most programs won't even use all of those. Storage benchmarks seem to actually suffer sometimes with Hyperthreading turned on.

With hyperthreading turned on the power usage is a huge jump up compared to without. Power pretty much stays under 150 watts with it turned off even when running AVX code but I have seen momentary jump to 240+ watts when running AVX code before system could react to bring it down under 200 watts when running with Hyperthreading. This was instantaneous & quickly remedied. Real life programs would likely never hit these power figures as there are wait times to load needed data in real programs.

I would like to say that the reason I did not list the specs of my system is that the spec would not have been needed to answer my question as I'm not trying to solve an issue. Just looking for the reason why something is the way it is. There was no actual problem to solve if my question was to be properly understood.
 
Last edited:

germanium

Distinguished
Jul 8, 2006
93
9
18,565
Found out that intel in their testing of Optane drives disabled hyperthreading.

https://www.intel.com/content/dam/w.../product-briefs/800p-series-product-brief.pdf

Read the number one reference at bottom of pdf.

By the way I just updated to 800p 118GB from the Optane Memory 32GB drives today, read speeds slightly higher, write speeds 2x+ faster. in my CPU attached RAID 0 X3 setup.

Sequential read 4350MB/s previous 4220MB/s slight improvement on Crystaldiscmark.
sequential write 1940MB/s previous 900MB/s 2x+ improvement on Crystaldiscmark.
Random read 4K QD1 235MB/s no change here on anvil storage benchmark.
Random Write 4K QD1 190MB/s previous 149MB/s good improvement here on anvil storage benchmark.
These are best results on each test type.

Latency slight improvement with RAID 0 compared to just 1 drive attached to CPU 203MB/s compared to 214MB/s reads on CPU attached RAID 0 on current window 10 1903 build. This build has much lower performance than the 1809 build in random reads Build 1809 314MB/s compared to 203MBs on lone drive on build 1903 on Crystaldiscmark, Other test programs not as dramatic reduction but still substantial. No other changes to my system. Had to install 1809 in order to setup CPU attached RAID then copy 1903 from my previous setup from a backup so did some testing while on 1809. all these tests were performed without write caching being enabled.
 
Last edited:

germanium

Distinguished
Jul 8, 2006
93
9
18,565
Did clean install of build 1903. Performance drop not due to 1903 as initially thought but due to cumulative update (kb4505057) which includes a previous cumulative update that has mitigations for new class of speculative execution side channel vulnerabilities. Unfortunately has large impact on Intel processors. This is especially true of small file storage latency. Up to 30% drop in performance at 4k que depth one on Crystaldiscmark. Other benchmarks not affected as much but still quite substantial.

Performance was same as 1809 build until this cumulative & previous update that had these mitigations.

Interesting to note as well that there is no latency penalty when using RAID 0 with CPU attached storage unlike with VROC, none at all. Performed exactly the same @ 4k que depth 1 with one drive as with 3 drives in RAID 0. Previous post indicating slight improvement is within the realm of test to test variation. There is a huge latency advantage to running CPU attached storage in RAID 0 compared to running any class of storage from the chipset.

https://support.microsoft.com/en-us/help/4505057/windows-10-update-kb4505057 which contains the following update.
https://support.microsoft.com/en-us/help/4497936
 
Last edited:

germanium

Distinguished
Jul 8, 2006
93
9
18,565
Did clean install of build 1903. Performance drop not due to 1903 as initially thought but due to cumulative update (kb4505057) which includes a previous cumulative update that has mitigations for new class of speculative execution side channel vulnerabilities. Unfortunately has large impact on Intel processors. This is especially true of small file storage latency. Up to 30% drop in performance at 4k que depth one on Crystaldiscmark. Other benchmarks not affected as much but still quite substantial.

Performance was same as 1809 build until this cumulative & previous update that had these mitigations.

Interestingly to note as well that there is no latency penalty when using RAID 0 with CPU attached storage unlike with VROC, none at all. Performed exactly the same @ 4k que depth 1 with one drive as with 3 drives in RAID 0. Previous post indicating slight improvement is within the realm of test to test variation. There is a huge latency advantage to running CPU attached storage in RAID 0 compared to running any class of storage from the chipset and there is no latency penalty at all for RAID 0 even with 3 drives.

https://support.microsoft.com/en-us/help/4505057/windows-10-update-kb4505057 which contains the following update.
https://support.microsoft.com/en-us/help/4497936
 
Last edited: