WTB: Motherboard 6driver raid5 support

[SvenBent]

So recently my driver in my file sever went belly up and i bought 4 new 3TB driver
but they are to big for my old intel Q45/Ich10r based chipset.
Trying to update the driver it get a BSOD ( this is an oem motherboard from a dell system

so im looking for a motherboard with the following specs
- Socket 775 or 1055 ( i have a i7 2400 as well
- 6 sata ports (4 able to do raid5) that support 3TB drives
- at least 1gbit ethernet (prefferably 2)
- Onboard VGA

any cheap offers ?

 

[Doozer]

bump because I looked. Good luck with your search!!

 

[SvenBent]

bumped

 

[Dan_D]

Any upper end Intel chipset based board would support RAID 5. However, I will tell you that the performance will be horrendous. You want a dedicated XOR processor for RAID 5, the chipset's "fake RAID" doesn't have that.

 

[SvenBent]

Any upper end Intel chipset based board would support RAID 5. However, I will tell you that the performance will be horrendous. You want a dedicated XOR processor for RAID 5, the chipset's "fake RAID" doesn't have that.

Any Core2duo or later will have no issues doing the Xoring ( it is the fastest instructions on a X86 cpu).
The bottleneck will still be the 1gbit network. XOR processing unit or not.

 

[Dan_D]

Any Core2duo or later will have no issues doing the Xoring ( it is the fastest instructions on a X86 cpu).
The bottleneck will still be the 1gbit network. XOR processing unit or not.

Incorrect. I did this on a Core 2 Quad system back in the day, and the performance was unacceptable. It was so bad I changed the configuration in short order.

 

[SvenBent]

Incorrect. I did this on a Core 2 Quad system back in the day, and the performance was unacceptable. It was so bad I changed the configuration in short order.

Well i apologize for you lack of abilities or what else hindered you in setting it up correctly
But I've ran this way for close to a decade with 115mb transfer speed on a Core 2quad.

You might have experience is due to the chunk alteration increasing writte amplication (kinda like on a SSD) but that is easily fixed but proper caching techniques and setting up the Intel Raid software correctly

In short: You rinstabilities to executes something successfully does not mean a limitation on other to do so. That's why we don't deal in negative burdens of proof.

 

[SvenBent]

Incorrect. I did this on a Core 2 Quad system back in the day, and the performance was unacceptable. It was so bad I changed the configuration in short order.

Just to ensure this is not something I just made up here is a post of my setup back from 2016
https://hardforum.com/threads/pre-built-nas-vs-diy.1918342/#post-1042670107

So again. Yes i absolut possible you just need to know how to set it up


P.S. Local transfer speed in Atto was peaking around 300MB/s way above 1gbit so yeah it was indeed the 1gbit nic making a bottleneck

 

[Dan_D]

Well i apologize for you lack of abilities or what else hindered you in setting it up correctly
But I've ran this way for close to a decade with 115mb transfer speed on a Core 2quad.

You might have experience is due to the chunk alteration increasing writte amplication (kinda like on a SSD) but that is easily fixed but proper caching techniques and setting up the Intel Raid software correctly

In short: You rinstabilities to executes something successfully does not mean a limitation on other to do so. That's why we don't deal in negative burdens of proof.

It's not hard to set up the software. I'm not sure what I could have done differently. I set up the arrays in the RAID BIOS of the motherboard. All of that predates any of the chipsets which later offered SSD caching for mechanical drive volumes. I've also dealt with the Intel RAID software hundreds of times. All I remember from the few times I've tried RAID 5 with onboard controllers, it was all Core 2 era or earlier, and it negatively impacted system performance. I'd get hesitation when accessing volumes. The actual transfer rates were fine. They were slower than what I saw using actual hardware controllers from LSI and Areca, but they weren't too bad.

That said, I've used Soft RAID in Linux even earlier and after that and it was fine. I don't know what to tell you. That was my experience. I've run RAID 5 setups on thousands of servers with dedicated hardware controllers without any issue. I've built hundreds of them personally. RAID 5 on a more modern Intel chipset is probably fine, given the CPU's we have today, but we are talking about Core 2 era here, which is the same level hardware I tried it on last and I wasn't pleased with the performance.

In that thread, the topic was NAS. I wasn't using it as a NAS. I was on the local system and it was sluggish accessing volumes. Had I been accessing the storage via network, I might not have noticed this. We are talking about two very different usage scenarios. When I was dissatisfied with the performance, I bought an LSI MegaRAID 8308ELP which cleared things right up. I did everything in hardware and it was plenty fast.

 

[SvenBent]

It's not hard to set up the software.

Yet you had issues with performance

I'm not sure what I could have done differently. I set up the arrays in the RAID BIOS of the motherboard. All of that predates any of the chipsets which later offered SSD caching for mechanical drive volumes. I've also dealt with the Intel RAID software hundreds of times. All I remember from the few times I've tried RAID 5 with onboard controllers, it was all Core 2 era or earlier, and it negatively impacted system performance. I'd get hesitation when accessing volumes. The actual transfer rates were fine. They were slower than what I saw using actual hardware controllers from LSI and Areca, but they weren't too bad

Correcl caching . not the defaults ones. and optimal stripe size is important factors ( more later)

That said, I've used Soft RAID in Linux even earlier and after that and it was fine. I don't know what to tell you. That was my experience. I've run RAID 5 setups on thousands of servers with dedicated hardware controllers without any issue. I've built hundreds of them personally. RAID 5 on a more modern Intel chipset is probably fine, given the CPU's we have today, but we are talking about Core 2 era here, which is the same level hardware I tried it on last and I wasn't pleased with the performance.

I have no opinion on comment to linux i havent worked with it

In that thread, the topic was NAS. I wasn't using it as a NAS. I was on the local system and it was sluggish accessing volumes. Had I been accessing the storage via network, I might not have noticed this. We are talking about two very different usage scenarios. When I was dissatisfied with the performance, I bought an LSI MegaRAID 8308ELP which cleared things right up. I did everything in hardware and it was plenty fast.

The thread was about NAS vs building yoursel. MY Post was on my fileserver and why i dont like prebuilt nas because their performance sucks compared to a simple C2Q setup with a intel raid controller


The issues with raid 5 is it has a shitton of write amplification.
Transfering files over windows fiel shares gives you chunck 8 kbytes i believei it is in samba.
When you have to write that 8KB ti a stripesie of lets say 64kb ( i belives this is the defaults) then you have to read the 64kbytes stripes size from each drive. modify one chunk with the new 8kbyts data (in memmory) recauclate the paritiy ( pretty fast) and then write back the 64kbytes chunks size to each disk
There is no way to calculatethe parity without knowing the data in the full stripe so you needs to read those data

on a 4 disk setup you 8 kbytes writes becomes a
256kb read and 256kbyts writes. That will KILL performance

The Intel raid setup manager is setup to bypass write cache to begin with. so these read/writes/ calcualtions has to be done in full before the disk will get out of bussy mode and accept the next 8bkytes of transfered data..
Adjusting the intel caching to include the write cache and disable the cache buffer flushing make you able to continues received transfered data and write it in "Burst"

so back to the 8kbyts samba delivery at a time
you can know queue up 16x 8kbyts writes in the cache. and now you only need to do the read 256kbytes and write 265kbytes data once for 16 transfer of 8bkytes
With the defautls cache you would have to read/write 16 times as much data.

If you can fill out the entire stripe in one transfer the controler does not need to read at all because it has the entire strip in memory and just needs to calcualte tha parity and write out the entire strip in one burst
now we are at no added I/O activity from raid 5 down from 32x write and 32x read amplification


Now i mostly transfer "big" files at least in the megabyte size so i am easily filling out entire stripe at at time keeping the read/write amplification down with the 64kbytes stripe size
but if you are doing smaller files you need to lower the stripe size to to a size that make you able to most of the time be transfering in full stripes.



Now im not saying that performance is NOT better with real XOR chip on the controller ( i used to have the old sx6000 controller)
but im still standing strong behind my arguments that for a simple fileserver (not database) a setup with a C2D and intel ichr controller will be more than enough to deliver full 1gbit transfer rates

 

[Beaflag VonRathburg]

I'm not reading these paragraphs that have nothing to do with the original post. If you want something that works and will be a minimal investment, purchase this: https://www.ebay.com/itm/Supermicro...JdJ34J:sc:USPSPriorityFlatRateBox!32225!US!-1 . The motherboard has a built in LSI 2308. Then, buy an i3 4130 or e3 1270 v3 (both are cheap and support ECC. It depends on what you need CPU wise), ECC UDIMM, sell all your current items, and move on with newer stuff. You'd be out of pocket about $100 for an i3 4130 and the motherboard before selling anything it replaces.

 

[Dan_D]

Yet you had issues with performance



Correcl caching . not the defaults ones. and optimal stripe size is important factors ( more later)



I have no opinion on comment to linux i havent worked with it


The thread was about NAS vs building yoursel. MY Post was on my fileserver and why i dont like prebuilt nas because their performance sucks compared to a simple C2Q setup with a intel raid controller


The issues with raid 5 is it has a shitton of write amplification.
Transfering files over windows fiel shares gives you chunck 8 kbytes i believei it is in samba.
When you have to write that 8KB ti a stripesie of lets say 64kb ( i belives this is the defaults) then you have to read the 64kbytes stripes size from each drive. modify one chunk with the new 8kbyts data (in memmory) recauclate the paritiy ( pretty fast) and then write back the 64kbytes chunks size to each disk
There is no way to calculatethe parity without knowing the data in the full stripe so you needs to read those data

on a 4 disk setup you 8 kbytes writes becomes a
256kb read and 256kbyts writes. That will KILL performance

The Intel raid setup manager is setup to bypass write cache to begin with. so these read/writes/ calcualtions has to be done in full before the disk will get out of bussy mode and accept the next 8bkytes of transfered data..
Adjusting the intel caching to include the write cache and disable the cache buffer flushing make you able to continues received transfered data and write it in "Burst"

so back to the 8kbyts samba delivery at a time
you can know queue up 16x 8kbyts writes in the cache. and now you only need to do the read 256kbytes and write 265kbytes data once for 16 transfer of 8bkytes
With the defautls cache you would have to read/write 16 times as much data.

If you can fill out the entire stripe in one transfer the controler does not need to read at all because it has the entire strip in memory and just needs to calcualte tha parity and write out the entire strip in one burst
now we are at no added I/O activity from raid 5 down from 32x write and 32x read amplification


Now i mostly transfer "big" files at least in the megabyte size so i am easily filling out entire stripe at at time keeping the read/write amplification down with the 64kbytes stripe size
but if you are doing smaller files you need to lower the stripe size to to a size that make you able to most of the time be transfering in full stripes.



Now im not saying that performance is NOT better with real XOR chip on the controller ( i used to have the old sx6000 controller)
but im still standing strong behind my arguments that for a simple fileserver (not database) a setup with a C2D and intel ichr controller will be more than enough to deliver full 1gbit transfer rates

As I said, it wasn't an issue with straight write performance as I recall. Keep in mind this was years ago when I last tried it. It was slowness accessing the volumes locally. If I ran a benchmark test, the results were what I expected and within decent values in line with what I saw others get online. I understand full well how the stripe size impacts these things. I deal with it every day on an enterprise level. I do not recall what I used for the stripe sizes on that configuration. I just remember clicking on a RAID 5 volume and it being incredibly slow to bring anything up. This wasn't an issue using the LSI MegaRAID adapter.

You may very well be correct. You've done it more recently than I have. I just recall it being so bad the couple of times I tried it myself that I never did it again. I also had a file server at work back in the day configured with a software RAID 5 and Windows NT that was atrocious. I didn't build that, it was like that when I got there. So, historically my experiences with it haven't been good on the Windows platform.

 

[SvenBent]

I'm not reading these paragraphs that have nothing to do with the original post. If you want something that works and will be a minimal investment, purchase this: https://www.ebay.com/itm/Supermicro...JdJ34J:sc:USPSPriorityFlatRateBox!32225!US!-1 . The motherboard has a built in LSI 2308. Then, buy an i3 4130 or e3 1270 v3 (both are cheap and support ECC. It depends on what you need CPU wise), ECC UDIMM, sell all your current items, and move on with newer stuff. You'd be out of pocket about $100 for an i3 4130 and the motherboard before selling anything it replaces.

Ty i'll look at it when i get back home. i cant check ebay at work
However i did find this one on newegg as well
https://www.newegg.com/p/1JW-0019-0..._-dell_-lenovo_-hp-_-9SIACTN9YR4413-_-Product
Ill compare at home
and this one
https://www.newegg.com/p/1JW-0006-0..._-dell_-lenovo_-hp-_-9SIACTN9YR4400-_-Product
Which according to forum post can take an I% 3rd gen, so i am assuming my current I5 2400 would fit fine in it

As I said, it wasn't an issue with straight write performance as I recall. Keep in mind this was years ago when I last tried it. It was slowness accessing the volumes locally. If I ran a benchmark test, the results were what I expected and within decent values in line with what I saw others get online. I understand full well how the stripe size impacts these things. I deal with it every day on an enterprise level. I do not recall what I used for the stripe sizes on that configuration. I just remember clicking on a RAID 5 volume and it being incredibly slow to bring anything up. This wasn't an issue using the LSI MegaRAID adapter.

You may very well be correct. You've done it more recently than I have. I just recall it being so bad the couple of times I tried it myself that I never did it again. I also had a file server at work back in the day configured with a software RAID 5 and Windows NT that was atrocious. I didn't build that, it was like that when I got there. So, historically my experiences with it haven't been good on the Windows platform.

I can fully admit that had i not made the changes i did on my file server. The Raid5 performance was tremendously bad the ICH5R.
But I have been running with raid 5 since Pata drive on the Abit BP6 motherboard
But still XOR calculation are seldom the underlaying issues as those are pretty fast on any cpu made within the last decade

Even programming on old 386 you would use Xr to reset a register because its faster to XOR a register against itself, than to write a 0 to it.

My ichr10r setup i would not run in and interprise due to the write cache. That is the big benefits in my eyes going with an LSI or spect controller. is that you can have a write cache with battery backup.
ICH10R with enabled write cache is not generel "best pratice" for enterprise enviroments