RAID 0 x2 vs a single drive - Gaming performance

[Ao1]

Set Up
· Target Array - RAID 0 128K Strip - Samsung 840 Pro 256 GB and Plextor M5 Pro 256 GB
· Target Drive – Samsung 840 Pro 256 GB
· OS Drive – Crucial M4 256 GB
· Steam folder located on target drives
· Intel RST Driver: 11.7.0.1013
· Graphics – AMD HD 6900 series
· CPU – i7-2700K @ 3.50GHz
· OS Windows 7

Monitoring Software
· hIOmon Disk I/O Ranger Version 7.0.600.0

Optimisations to help reduce I/O unrelated to tasks being monitored:
· AV disabled
· Indexing Disabled
· Page file off
· System Restore off

Tasks Monitored (All tasks are monitored at the physical device level)

· Crysis 2 – In at the deep end level
· Crysis 2 – All levels loaded
· Black Ops 2 – Pyrrhic Victory

Before looking at gaming performance I ran a number of sequential I/O access patterns using Iometer on a logical drive with one worker at queue depth one and monitored performance with the Disk I/O Ranger. The results from Iometer and the Disk I/O ranger are similar (measurements between Iometer and the Disk I/O Ranger are taken at different points in the I/O stack so they vary slightly). This provides a benchmark for the theoretical performance of the data transfer sizes monitored.


Uploaded with ImageShack.us

Data Transferred/Time Index (DXTI)
The DXTI provides a rating of overall performance efficiency. Higher is better. In the table below it can be seen that the RAID 0 array is faster than the single drive for the three tasks that have been monitored. For the data xfers sizes monitored the workload primarily consisted of 100% sequential I/O operations for data xfers sizes below 256 KiB. Data xfers sizes above 256 KiB were between 50% and 100% random.


Uploaded with ImageShack.us

Crysis 2 – In at the deep end level
In the chart below the Max Observed MiB/s for the monitored data xfers sizes are compared. The RAID 0 array does not benefit from faster MiB/s speeds for data xfers sizes above 128 KiB and the MiB/s speed for data xfers sizes 512 KiB and above are identical. MiB/s speeds are however generally faster for data xfers sizes below 128 KiB.


Uploaded with ImageShack.us


In the chart below a max queue depth comparison is provided between the RAID 0 array and the single drive. There is not a significant difference between either storage set up.


Uploaded with ImageShack.us

In the table below the percentage of I/O operations with a response time < 1ms (fast I/O) are compared. The RAID 0 array has a higher percentage of fast I/O’s, indicating higher efficiency at transferring data.


Uploaded with ImageShack.us

In the table below the percentage of I/O operations within response times ranges are compared.


Uploaded with ImageShack.us

In the table below the amount of data transferred within the response time ranges is compared. Overall the RAID 0 array is able to move more data with lower response times.


Uploaded with ImageShack.us

In conclusion the additional bandwidth capability of the RAID 0 array does not benefit this task, but (somewhat surprisingly) the RAID 0 array does benefit from reduced response times, especially in the RT<50 us range.

Crysis 2 – All Levels
In the table below I compare MiB’s performance over all levels of Crysis 2. (Load level, exit, load next level, exit etc). There is not a significant difference between the results from loading the “In at the deep end” level.


Uploaded with ImageShack.us


Black Ops 2 - Black Ops 2 – Pyrrhic Victory
Once again the RAID 0 array is able to achieve a higher percentage of fast I/O’s


Uploaded with ImageShack.us


In the table below the single drive is actually faster for 512 KiB & 1 MiB xfers sizes. Bandwidth capability of the RAID 0 array is again not utilised.



Uploaded with ImageShack.us

 

[Ao1]

Continued...

In the tables below it can be seen that the RAID array is overall more efficient at transferring more data within lower response time ranges.



Uploaded with ImageShack.us


Uploaded with ImageShack.us

Summary
Quiet surprisingly the increased bandwidth capability of RAID 0 array did not bring any additional benefit over a single drive, however the RAID 0 array was able to reduce response times, providing faster overall I/O performance. The workload predominantly consisted of small transfer sizes that were sequential. The Samsung 840 Pro and Plextor M5Pro both benefit from device re-ahead caching if the workload is sequential and perhaps it is the combined read ahead cache capability of each drive within the array that enabled the faster response times.

 

[wpcoe]

What tables?

(EDIT: Possibly a fluke of forum weirdness, but when I first posted there were NO tables in the OP! Honest!)

 

[Ao1]

Below are the screenshots taken for Crysis 2 – In at the deep end level, which provided the data for the graphs. There is a lot of info on the screenshots. Too much to explain, but details of what the metrics relate to can be found here if of interest:

http://www.hyperio.com/hIOmon/Documentation/hIOmonDIORUserGuide.pdf

Single Drive Results


Uploaded with ImageShack.us

RAID 0 Rsults


Uploaded with ImageShack.us

 

[groberts101]

Nice writeup there and totally expected for that type of data flow, IMO.

Just out of curiosity.. did you enable write-back caching on the array first?

I can only assume that someone as knowledgable as yourself must realize that it is off by default on all Intel based raids.

I also have to wonder what changes would occur with all data located on the array versus splitting it off to the target drive like that(although the sata chip shouldn't bottleneck as it has more than sufficient bandwidth to share evertyhing back and forth without issue).

Not that it's uncommon to do it the way you have the system configured there.. but I do know that system with a single drive setup running all files on the same drive can get bigger boosts when running that same configuration(single volume systems) when running raids.

 

[kbp]

Now change that stripe size to 32K or even 16K and watch what happens.

 

[Ao1]

A 32K strip makes very little difference. There is a very small reduction in response times overall and 4 KiB performance is improved.


Uploaded with ImageShack.us

 

[jwilliams4200]

Could you post some more information about your system?

1) Motherboard make and model

2) Screenshot of the Windows device manager Policies tab for the "disk drives" you tested (showing write cache and buffer flushing settings) -- include both the single SSD and the RAID

3) Screenshot of the Intel RST application, showing the "Manage Disk" (with advanced tab opened) for the single-SSD and the RAID that you tested

Also, it would be helpful to see an ASU (Anvil Storage Utilities) benchmark screenshot for the single SSD and for the RAID.

 

[Ao1]

RAID 0
· Write Caching on the Device &#8211; Enabled
· Windows write cache buffer flushing &#8211; Off
· RST Write-back cache &#8211; Enabled

Single Drive
· Write Caching on the Device &#8211; Enabled
· Windows write cache buffer flushing &#8211; On

Mobo/ RAM details on the ASU Screenshots. No doubt it would be faster if I was using two 840 Pro&#8217;s.

I tried disabling RST write-back cache and it made no difference when I re-ran the Crysis 2 &#8211; &#8220;In at the deep end level&#8221;. When device write cache is disabled in RAID 0, or on a drive set as a non RAID member disk, seq read performance does not change significantly on an Iometer run. I need to double check but I&#8217;m sure that disabling device write cache has an impact on sequential read performance on a single drive in AHCI mode. (At least it does on the M5Pro and 830. I haven&#8217;t checked the 840Pro yet).

840 Pro (Non RAID Member)


Uploaded with ImageShack.us

RAID 0 128K Strip


Uploaded with ImageShack.us

RAID 0 32K Strip


Uploaded with ImageShack.us

 

[christer12]

If I overclock my PC, has it any effect on SSD hard drive performance?

 

[jwilliams4200]

RAID 0
· Write Caching on the Device – Enabled
· Windows write cache buffer flushing – Off
· RST Write-back cache – Enabled

That is ambiguous (trying to avoid ambiguity is why I reqeusted a screenshot of the Policies tab).

When you say that "write cache buffer flushing" is "Off", do you mean that you have a check mark in the box next to "Turn off Windows write-cache buffer flushing on the device"? That would be the literal interpretation of what you wrote. Or do you mean you have an empty (unchecked) box?

 

[Ao1]

I got a new mobo (Asrock Z77 Xtreme11) which has an onboard LSI controller supporting up to 8 drives via a PCI 3.0 x8 connection. In the manual it states that PCIe 3.0 for GPU is only supported with Ivy Bridge, so maybe that is why I can only get ~2.6GB/s on the LSI chip, as I'm using Sandy Bridge. The advertised max bandwidth is 3.8GB/s.

The LSI chip does not appear to scale at higher QD's either. Overall a big disappointment, although maybe things would improve with an Ivy Bridge CPU.

Anyway for what it's worth here is a comparison of Crysis using RAID 0 x8 (8 x 256GB SSD's)

The max MiBs observed for larger data transfer sizes (128 KiB and >) remain almost identical.


Uploaded with ImageShack.us

The percentage of I/O operations <50 us is significantly below the Intel RAID 0 x2 array and single drive (840Pro), however less I/O's are occurring with higher rsp times. (The onboard LSI controller has no cache)


Uploaded with ImageShack.us

Uploaded with ImageShack.us

There aren't many configuration options available via LSI MagaRAID storage manager. The only configurable option was the disk cache policy.


Uploaded with ImageShack.us

Uploaded with ImageShack.us


Uploaded with ImageShack.us

 

[Ao1]

Here is a summary of DXTI scores with various storage configurations


Uploaded with ImageShack.us

 

[Ao1]

Crysis 3
Set Up

Source Drive - Plextor M5 Pro 256 GB
OS Drive – Samsung 840 Pro 256 GB
Intel RST Driver: 11.7.0.1013
Graphics – AMD HD 6900 series
CPU – i7-2700K @ 3.50GHz
OS - Windows 8 64bit
Mobo - Asrock Xtreme11 Z77

Key Stats

· Data Read – 2,392 MiB
· Number of read I/O operations = 28,529
· % of random read I/O operations = 50.07%
· % of data transferred via random read I/O = 33.71%
· % of read I/O operations below 1ms = 98.85%
· % of read I/O operation above queue depth 1 – 1.95%
· Max response time = 29.64 ms
· Avg Response time = 0.401 ms
· Max IOPS of data transfer sizes (DTS) monitored = 266 (128KiB)
· 4 KIB max IOPS = 130
· Max MBs = 65.01 (DTS 2 MiB)


Uploaded with ImageShack.us

 

[tweakboy]

I had a friend who did this one time and wasnt too pleased with results. He told me its the same s*it . But benchmarks are very high. gl

 

[Ao1]

Crysis 3 &#8211; 40 Minutes Single Player
Key Stats

·Data Read &#8211; 10,591 MiB
·Number of read I/O operations = 138,834
·% of random read I/O operations = 49.24%
·% of data transferred via random read I/O = 33.13%
·% of read I/O operations below 1ms = 99.52%
·% of read I/O operation above queue depth 1 &#8211; 1.37%
·Max response time = 644.563 ms
·Avg Response time = 0.376 ms
·Max MBs = 53.87(DTS 128 KiB)



Uploaded with ImageShack.us


Battlefield Bad Company 2 Single Player
·Data Read &#8211; 454.62 MiB
·Number of read I/O operations = 2,708
·% of random read I/O operations = 19.68%
·% of data transferred via random read I/O = 8.34%
·% of read I/O operations below 1ms = 98.44%
·% of read I/O operation above queue depth 1 &#8211; 12.56%
·Max response time = 20.476 ms
·Avg Response time = 0.502 ms
·Max MBs = 64.75(DTS 256 KiB)



Uploaded with ImageShack.us

 

[Ao1]

Black Ops 2 &#8211; 40 Minutes Multi Player
Key Stats

·Data Read &#8211; 1,972.39 MiB
·Number of read I/O operations = 13,288
·% of random read I/O operations = 60.35%
·% of data transferred via random read I/O = 44.29%
·% of read I/O operations below 1ms = 93.12%
·% of read I/O operation above queue depth 1 &#8211; 11.42%
·Max response time = 20.839 ms
·Avg Response time = 0.571 ms
·Max MBs = 81.79 (DTS 256 KiB)


Uploaded with ImageShack.us

 

[Ao1]

Snipe Ghost Two – 1 hour 35 Minutes Single Player
Key Stats
· Data Read – 1537.11 MiB
· Number of read I/O operations = 10,562
· % of random read I/O operations = 45.81%
· % of data transferred via random read I/O = 28.65%
· % of read I/O operations below 1ms = 84.54%
· % of read I/O operation above queue depth 1 – 12.54%
· Max response time = 23.371 ms
· Avg Response time = 0.617 ms
· Max MBs of monitored DTS = 33.55 (DTS 2 MiB)
· Max IOPS of monitored DTS = 420 (4KiB)



Uploaded with ImageShack.us

 

[Ao1]

A quick look at writes on the OS drive (DR0) over a ~16 hour duration

· Data Written &#8211; 1,163.14 MiB
· Number of write I/O operations = 33,326
· % of random write I/O operations = 90.16%
· % of data transferred via random write I/O = 72.78%
· % of write I/O operations below 1ms = 92.83%
· % of write I/O operation above queue depth 1 &#8211; 22.81%
· Max response time = 809.55 ms
· Avg Response time = 2.46 ms
· Max MBs of monitored DTS = 124.78 (DTS 1 MiB)
· Max IOPS of monitored DTS = 477 (4KiB)
· Max MBs of DTS 4KiB = 1.95
· 54% of write I/O&#8217;s were based on a 4 KiB DTS
· 43% of write I/O&#8217;s were undertaken with a response time of less than 50us, but the average and maximum response times are much higher than for read operations.
· 14 write I/O&#8217;s had a response time of 500ms and above
· 9 write I/O&#8217;s occurred with a queue depth of 64 and above



Uploaded with ImageShack.us