Using hidden RAM in Windows Server 2008

[Axel Mertes]

I was looking for potentially the LSI controller or its successor tested here:

http://www.tomshardware.de/RAID_Control ... 98-11.html

They reach a performance that is fully in our aims with this. As we will probably use "standart" SATA SSDs instead of more expensive server style SAS ones, we plan to use RAID 5 or 6 for the forseeable event of damage of an SSD. I would think that in the long run it may turn out cheaper this way. From two dozen SSDs we run so far, we had just one failure, and that was a mechanical connectore issue. So however strong we used them yet (as caching drives), they hold up for years now. I would assume that after the warranty time is over, the next generation will be around the corner too, so we may swap drives or controller at some point. But again, to feed a 20 GBit ethernet link, a 2+ GByte/s SSD RAID should be a good fit. There is very small latency and using RAM as first level cache would making the server firing data at yet unseen speed. Response times will be very small compared to what we experience now (I hope).

As disks we aim at e.g. the Samsung 840 EVO 1000 GByte or the Crucial M550 1000 GByte ones. Given that each one of those is rated around 500 MByte/s, we should be fairly safe using 8 of them to get around 2 GByte/s total.

When writing data, I would expect that system is first using local RAM as write cache to collect for bigger writes. Then write them through to normal HDD storage for safety reasons (not loosing anything in the case of a power failure of any kind). From the RAM cache data should be written (deferred) to SSD cache to stay in a fast cache media of later re-use (again rendering & review is most of the tasks). These deferred writes should minimize the SSD write cycles.

Is kind of a tiered storage model, feeding data from caches first, then HDDs. It works this transparent in the server. My wish would be to implement this architecture in every clients (so server and client sides). However, block storage and SMB shares aren't the same game, and a SAN sharing requires a global cache control (in terms of invalidating cached data blocks which have been changed by another machine)

What do you think?

[Violator]

I was looking for potentially the LSI controller or its successor tested here:

http://www.tomshardware.de/RAID_Control ... 98-11.html

They reach a performance that is fully in our aims with this. As we will probably use "standart" SATA SSDs instead of more expensive server style SAS ones, we plan to use RAID 5 or 6 for the forseeable event of damage of an SSD. I would think that in the long run it may turn out cheaper this way. From two dozen SSDs we run so far, we had just one failure, and that was a mechanical connectore issue. So however strong we used them yet (as caching drives), they hold up for years now. I would assume that after the warranty time is over, the next generation will be around the corner too, so we may swap drives or controller at some point. But again, to feed a 20 GBit ethernet link, a 2+ GByte/s SSD RAID should be a good fit. There is very small latency and using RAM as first level cache would making the server firing data at yet unseen speed. Response times will be very small compared to what we experience now (I hope).

As disks we aim at e.g. the Samsung 840 EVO 1000 GByte or the Crucial M550 1000 GByte ones. Given that each one of those is rated around 500 MByte/s, we should be fairly safe using 8 of them to get around 2 GByte/s total.

When writing data, I would expect that system is first using local RAM as write cache to collect for bigger writes. Then write them through to normal HDD storage for safety reasons (not loosing anything in the case of a power failure of any kind). From the RAM cache data should be written (deferred) to SSD cache to stay in a fast cache media of later re-use (again rendering & review is most of the tasks). These deferred writes should minimize the SSD write cycles.

Is kind of a tiered storage model, feeding data from caches first, then HDDs. It works this transparent in the server. My wish would be to implement this architecture in every clients (so server and client sides). However, block storage and SMB shares aren't the same game, and a SAN sharing requires a global cache control (in terms of invalidating cached data blocks which have been changed by another machine)

What do you think?

I would go for the Crucial drives, those are rebranded Mikron drives and also way more secure to use than the TLC based Samsung Evo's.
Personally I would get something like a HP Lefthand setup or aim higher with a 3Par solution with a 4 business hours support plan added, higher cost but worth it if your company can cover it without problems.

Your write/read strategy makes sense, only need that fully automated with a smart release mechanism, last part will be a bit tricky, since you most probably want to ensure that each user has the cache which is in works cached, and the rest flushed out.

I would go for 2012R2 instead of 2008R2, especially when you look at how fast SMB 3.0 is, that depends if the rest of your hard and software is up to date to work with it.

[Axel Mertes]

Interesting note about the Crucial drives. Never heard that before. Thanks!

As a small company something like HP Lefthand and 3Par are already a bit out of budget. Just had a dicounted offer last week and it was about the entire annual budget I get for IT

I read often about 2012R2 being a lot better in SMB than 2008R2. I haven't had the chance yet to test myself, however, it sounds like a serious improvement worth the investment.
Do you know if we need CALs when going for 2012R2 Standard edition?
With 2008R2 we did not need them at all, as we don't use any active directory / domain login stuff, just peer to peer pure network SMB server. If I got it right, that was changed in 2012R2. Also, I don't need hidden memory then... (what a waste in 2008R2 Standard...)

Most clients are Windows 7 and a very few Windows 8 (the latter are only notebooks, so far). Would that work with SMB 3.0?
It seems there is no SMB 3.0 implementation for Windows 7 unfortunately. However, as we aim at single 10 GBit links to clients for the beginning, it should't make much of a difference. We need to check if all apps and hardware finds Windows 8 drivers...

[Violator]

Interesting note about the Crucial drives. Never heard that before. Thanks!

As a small company something like HP Lefthand and 3Par are already a bit out of budget. Just had a dicounted offer last week and it was about the entire annual budget I get for IT

I read often about 2012R2 being a lot better in SMB than 2008R2. I haven't had the chance yet to test myself, however, it sounds like a serious improvement worth the investment.
Do you know if we need CALs when going for 2012R2 Standard edition?
With 2008R2 we did not need them at all, as we don't use any active directory / domain login stuff, just peer to peer pure network SMB server. If I got it right, that was changed in 2012R2. Also, I don't need hidden memory then... (what a waste in 2008R2 Standard...)

Most clients are Windows 7 and a very few Windows 8 (the latter are only notebooks, so far). Would that work with SMB 3.0?
It seems there is no SMB 3.0 implementation for Windows 7 unfortunately. However, as we aim at single 10 GBit links to clients for the beginning, it should't make much of a difference. We need to check if all apps and hardware finds Windows 8 drivers...

Your welcome, we bought some desktops for our web developers which got these drives, so had to figure out what exactly these Mikron SSD's are all about, and found out that Crucial sells them as rebranded.

You need CAL's for Standard per 2 CPU sockets, Data Center requires one per 2 CPU sockets as well, but no point in that unless you virtualize with Hyper V, since you afterwards can install whatever version of 2012R2 you want as VM without purchasing additional licenses, or if you need more RAM available.
http://www.microsoft.com/licensing/abou ... aspx#tab=2

We did some testing with and SMB 3.02 / Hyper-V environment to see how it performs compared to Block Level / VM-Ware (plus the feature differences between them).
Live migration was really faston the crappy test setup we used, I mean, 2 minutes to move a 2012R2 VM while live streaming full HD on that same VM without interruption..
All Hyper-V machines used WD Greens and 4 not so good performing SSD's striped up as NAS, connected to some old Cisco 100Mbit switches (we had 4x100Mbit bundled up as Ether Channel though besides 4 ports on each PC, 1 for management and 3 for load balancing and failover).
I would love to see how Branche Cache performs with SMB 3.0, I can only imagine that it would be way faster.
Point is, SMB 3.0 is way faster than MS iSCSI, 3rd party iSCSI is another matter since most of them provide multi channel iSCSI ( MS does not) which can be faster than SMB 3.0, SMB 3.0 can handle 10Gps though.
3rth party ads additional cost though, so if your budget is tight it might be worth considering, depends on how cheap you can build a solution with performance on pair.
The minus part is, that SMB 3.0 takes a bit more to setup before it is running, but that is half the fun of it.
You can evaluate it for 180 days, so possibility to test it out before setting anything live is there: http://technet.microsoft.com/en-us/eval ... 05286.aspx

As for Windows 7, they will communicate with SMB 2.1, so you would not get any performance gain for these clients, so no multi channel advantage.