Aggregate HD speeds capped at 150MB/s?

Childs · 02-04-2013, 01:08 PM

I got an SSD, and while benching the drive I started to bench the other drives. I have 2 x 1TB SATA2 drives in software raid 0 (D), and 1 1TB drive (C) that has the OS and apps, etc. This is on a Sabertooth X58, using the Intel SATA2 ports. Bios is set to AHCI, and I'm using the MS AHCI driver. Jmicron and Marvell controllers are disabled.

Benching the drives individually, D gets around 125MB/s. C gets 115MB/s. Now, using perfmon I monitored the transfer rates when copying between drives, they top out at 145-150MB/s total. 75MB/s read, and the other drive does 75MB/s write.

So my understanding is that the bandwidth isnt shared, so each port should get close to 280-300MB/s, and the SSD does on the same controller. So copying between any two drives should be at around 115MB/s per drive, or around 230MB/s aggregate. Its almost like the mechanical hard drives are using a SATA1 controller instead of SATA2. Can anyone think of what might be going on?

dave_the_nerd · 02-04-2013, 01:57 PM

Benching is one thing, file transfers are another. (file system overhead, small files vs. big files, disk fragmentation, etc.)

Write is generally lower than read.

Copying from drive A to drive B, you will be limited to the slowest speed of the slower drive, it's not aggregate.

RAID will add some CPU overhead too.

75 MB/sec is reasonable.

Childs · 02-04-2013, 02:11 PM

Even on a single big file (4GB iso) from a single mechanical HD or raid 0 HD to the SSD is still 75MB/s. It seems like there is some limit (150MB/s total, 75MB/s each drive) when two drives are in use at the same time. I actually have another SSD coming this week, but I want to know if this is some limitation on the chipset in my board. If its still going to be 75MB/s max while copying, I want to re-evaluate the upgrades I'm doing.

dave_the_nerd · 02-04-2013, 03:16 PM

What if you use the Intel chipset drivers instead of the MS AHCI ones?

imagoon · 02-04-2013, 03:22 PM

Any reason why you are not using the Intel AHCI drivers? The MS AHCI driver doesn't turn on everything on the new Intel chips. Try installing the RST.

Childs · 02-04-2013, 03:48 PM

Quote:

Originally Posted by dave_the_nerd

What if you use the Intel chipset drivers instead of the MS AHCI ones?

Quote:

Originally Posted by imagoon

Any reason why you are not using the Intel AHCI drivers? The MS AHCI driver doesn't turn on everything on the new Intel chips. Try installing the RST.

I used the Intel Drivers, and it didnt make a difference. In retrospect, I grabbed them from the Asus site and not from Intel. They might have been a year old. I reverted back to the MS driver when I was trying to isolate if the speeds issues were caused by the Jmicron esata controller or Marvell 6Gbps controller drivers. I read a random post somewhere saying that the MS AHCI driver might be better when you have these controllers all together.

henryay · 02-04-2013, 04:06 PM

Make sure your BIOS settings are set correctly.
Also, what motherboard/chipset do you have? There were some Marvell controllers released a few years ago that weren't so good. I have a P55 motherboard with a Marvell SATA3 controller and I had to mess with the BIOS settings to get over the 150MB/s cap. Because, the controller is not so good, I cannot get my SSD to its rated spec of 520MB/s. It is currently running at around 350MB/s.

Edit: Just saw you have the Sabertooth X58 and the Marvell controller disabled...

Childs · 02-04-2013, 04:20 PM

Quote:

Originally Posted by henryay

Make sure your BIOS settings are set correctly.
Also, what motherboard/chipset do you have? There were some Marvell controllers released a few years ago that weren't so good. I have a P55 motherboard with a Marvell SATA3 controller and I had to mess with the BIOS settings to get over the 150MB/s cap. Because, the controller is not so good, I cannot get my SSD to its rated spec of 520MB/s. It is currently running at around 350MB/s.

Edit: Just saw you have the Sabertooth X58. It uses the same Marvell controller that I have.

It has 3 sata controllers, a Intel ICH10R controller, Marvell 9128 and Jmicron JMB632. I have the Marvell and Jmicron disabled now. I'm not sure what else there is to check in the bios, as its set to AHCI. It does seem like there is some setting capping the speed, but I have no idea what to check.

groberts101 · 02-04-2013, 08:56 PM

try setting the raid up in the bios rather than using Windows software raid... and then retest.

Childs · 02-04-2013, 11:40 PM

I ran another test by creating a 50GB file with random data and copied it from my HD to the SSD, and the transfers are around 100-110MB/s reads at times, while doing about the same or more in writes to the SSD. I think the difference is the disk I am now reading from is only at 35% capacity, where as before the raid 0 drive was over 95% full. When I got rid of the raid volume and used another single disk, it was also at 95% full because I backed up the raid volume to that disk. So I always had a near capacity drive in the copying tests. At least now I see over 100MB/s or more per drive at times. What threw me off was benchmarking the individual drives and seeing over 100MB/s.

Anyways, thanks for everyone's input. I guess this means I dont need a new system. I had a new mb, cpu and ram in my shopping cart before I ran the last set of tests.

sub.mesa · 02-08-2013, 01:09 PM

Post your benchmarks. Or at least tell us what benchmark you used (HDtune maybe? that one is not suitable for RAID-arrays).

Childs · 02-08-2013, 01:20 PM

Quote:

Originally Posted by sub.mesa

Post your benchmarks. Or at least tell us what benchmark you used (HDtune maybe? that one is not suitable for RAID-arrays).

I was using Perf Mon to monitor realtime transfer rates between two disks on the same controller. I used Atto and CrystalDiskMark to test the drives individually.

sub.mesa · 02-08-2013, 01:33 PM

Alright. Well, without benchmarks I cannot say anything with authority. But from what you told me so far, everything is working as it should.

SATA/300 delivers about 250MB/s of usable throughput per harddrive (or SSD). But harddrives are not fast enough to saturate SATA/300. They do about 50MB/s - 150MB/s depending on the location you are testing, as well as the harddrive model.

Quote:

So copying between any two drives should be at around 115MB/s per drive, or around 230MB/s aggregate

This would be about right, assuming you are testing the fastest regions of a harddrive, meaning LBA close to zero. In HDtune, you can clearly see the sequential performance degrading as the harddrive approaches the end of its capacity. This is normal, since the inner tracks on a harddrive platter can contain less information per cycle than the outer tracks which per revolution would encompass much more media 'surface'.

So the real question is, how full were the drives. If you are testing drives which are more than a few percent full, you will already start to notice sequential transfers are a little bit slower. If you are testing a drive that is almost full, the maximum sequential transfer rate would be about half that of its maximum performance when the drive is empty.

Can you confirm to me whether both drives are filled up with existing data? If yes, the above should explain your issue.

Childs · 02-08-2013, 01:45 PM

Quote:

Originally Posted by sub.mesa

Alright. Well, without benchmarks I cannot say anything with authority. But from what you told me so far, everything is working as it should.

SATA/300 delivers about 250MB/s of usable throughput per harddrive (or SSD). But harddrives are not fast enough to saturate SATA/300. They do about 50MB/s - 150MB/s depending on the location you are testing, as well as the harddrive model.

This would be about right, assuming you are testing the fastest regions of a harddrive, meaning LBA close to zero. In HDtune, you can clearly see the sequential performance degrading as the harddrive approaches the end of its capacity. This is normal, since the inner tracks on a harddrive platter can contain less information per cycle than the outer tracks which per revolution would encompass much more media 'surface'.

So the real question is, how full were the drives. If you are testing drives which are more than a few percent full, you will already start to notice sequential transfers are a little bit slower. If you are testing a drive that is almost full, the maximum sequential transfer rate would be about half that of its maximum performance when the drive is empty.

Can you confirm to me whether both drives are filled up with existing data? If yes, the above should explain your issue.

Oh, I came to this conclusion on the 10th post. When I was initially did the testing at least one drive was 90%+ filled. I wasnt sure if one of the raid 0 disks was failing, so I cloned it to another bigger, faster drive, but again, it now made that drive 90% full. When I went to a 35% full drive, the speeds were at least 100MB/s per drive, 200MB/s+ aggregate on the controller. What threw me was the numbers from testing the drives individually, because they would be over 100MB/s even though they were almost full. I guess a sequential write on even the slower part of the drive can be sustained during a benchmark.