dslreports logo
 
    All Forums Hot Topics Gallery
spc
uniqs
20

koitsu
MVM
join:2002-07-16
Mountain View, CA
Humax BGW320-500

koitsu to aurgathor

MVM

to aurgathor

Re: Looking For SATA Controller Card For Server 2012 Essentials

Actually the storage card manufacturer can't add circuitry for the most part -- it would require Marvell either a) increasing the PCIe lane count the IC offers, or b) moving up to PCIe 3.0. Chances are they'd go with (a) and the chip would have a different model number.

The below assumes assumes you have a PCIe 2.0 slot that provides PCIe 2.0 x2 (or more) lanes, and not a PCIe 2.0 slot that supports x2 (or more) cards but only offers, say, x1 lanes. Anyway, to work out the math on this for those wondering:

PCIe 2.0 x1 = 1GByte/sec (1024MByte/sec)
PCIe 2.0 x2 = 2GByte/sec (2048MByte/sec)

SATA300 = supports up to 300MByte/sec
SATA600 = supports up to 600MByte/sec

(I'm also assuming a unit of 1024, not 1000 -- throughput, at least on networks, uses a unit of 1000, but I'll stick with 1024)

Standard MHDDs can do about 150-180MBytes/sec sequential from the platters -- note this does not reflect reads from the disk that are being fed by the on-disk cache (those can sometimes reach the SATA PHY speed, but I've rarely seen this with MHDDs).

So let's say you have 4x MHDDs that are magical and can do 180MByte/sec constantly (from LBA 0 to end of drive), sequential or random. I said magical, right?

4 * 180MB/s = 720MBytes/sec total

A single PCIe 2.0 x1 lane can handle that, with ~304MB bandwidth left over (say for cached reads). I think this would be fine.

The situation changes dramatically if a person decides to use present-day SSDs (ex. Samsung 840 series) with that controller, and the PCIe lane count becomes a serious bottleneck.

Since I have a Samsung 840 256GB SSD, and know what my sequential numbers are, let's use it as an example: 560MBytes/sec read, 256MByte/sec write.

Let's say you have 4 of those and stick them on the aforementioned controller:

4 * 560MB/s = 2240MBytes/sec read total
4 * 256MB/s = 1024MBytes/sec write total

So for reads you would be hitting a bottleneck capacity and losing, potentially ~192MByte/sec worth of throughput. For writes, no impact.

Thus -- the above controllers will do you just fine as long as you're using them predominantly for MHDDs. A mix of SSDs and MHDDs is fine too, just do the math and work things out. And don't forget about things like driver overhead and OS overhead too -- those numbers above are all just math, they don't reflect real-world numbers nor have I done any actual testing with such a controller to see what its limits are. But I do think using one of those for MHDDs exclusively would be perfectly fine.

aurgathor
join:2002-12-01
Lynnwood, WA

aurgathor

Member

By "circuitry" I meant an interface between an x4 PCI-e bus and the x2 inputs of the Marvell chip.

SATA III add in cards make most sense in older PCs that don't have SATA III, and those PCs usually only have PCI-e 1.0 slots.
JoelC707
Premium Member
join:2002-07-09
Lanett, AL

JoelC707

Premium Member

Yeah but what would the extra two lanes connect to? If the Marvell chip can only support x2, I guess they could add a second one and do 2 ports on one chip (it'd be an 8 port capable card with only4 active ports). I don't see them really doing that personally.

That said, I do mostly agree with who would be buying these cards, people that want SATA III but don't have them. People like Octavean See Profile are probably the exception and not the rule as for who would be using these. Chances are their secondary slots are NOT 2.0 compatible (though they may be) which effectively halves the numbers koitsu See Profile mentions and means there would be a bottleneck with an x2 card but not with an x4 card.

koitsu
MVM
join:2002-07-16
Mountain View, CA
Humax BGW320-500

koitsu to aurgathor

MVM

to aurgathor
said by aurgathor:

By "circuitry" I meant an interface between an x4 PCI-e bus and the x2 inputs of the Marvell chip.

The 88SE9235 only has a single PCIe 2.0 x2 lane though. To benefit from a full x4 lane, they'd need to either increase the ICs lane size to x4 or use (or possibly multiplex) two x2 lanes (coming from the same chip).

I guess alternately (and more costly), a card manufacturer could actually put two 88SE9235's on one card, and use only 2 SATA ports per chip (rather than 4) to help deal with the limited bus bandwidth per chip.
said by aurgathor:

SATA III add in cards make most sense in older PCs that don't have SATA III, and those PCs usually only have PCI-e 1.0 slots.[

I dunno, I don't exactly agree with that. I think any kind of SATA300 or SATA600 HBA has its use even on a present-day PC; today's systems will usually give you (2) SATA600 ports and (2) or (4) SATA300 ports. Many people actually max those out; I can't tell you how many times I've seen people complain about only having two SATA600 ports (i.e. they want to make a system with a bunch of SSDs). This is exacerbated by by the fact that many of those mainboard vendors add a 2nd chip (Marvell, JMicron, etc.) for additional SATA ports, and those are usually the chips that are crap / have crappy drivers / don't play well with everything else on the system.

I've been saying this for years, but I really wish Intel would make dedicated SATA/SAS ICs. Literally just take the SATA (and for some, SAS) logic out of their southbridges, put them into a dedicated IC and sell that to vendors -- or make their own HBA that uses that. And only two chip versions: one with RAID, one without. They could do 4 or even 6-port cards this way, honestly. I swear, for small or mid-size businesses these would sell like hotcakes for people doing storage, solely because of how reliable their existing stuff is. As it stands right now, I can only trust a handful -- if that! -- of after-market SATA HBA vendors.
Thordrune
Premium Member
join:2005-08-03
Lakeport, CA

Thordrune to koitsu

Premium Member

to koitsu
said by koitsu:

Anyway, to work out the math on this for those wondering:

PCIe 2.0 x1 = 1GByte/sec (1024MByte/sec)
PCIe 2.0 x2 = 2GByte/sec (2048MByte/sec)

IIRC, PCIe 2.0 is half of that bandwidth, per direction and lane. 5 GT/sec and 8b/10b encoding.
JoelC707
Premium Member
join:2002-07-09
Lanett, AL

JoelC707

Premium Member

Depends. If he is calculating transmit and receive together then those figures are correct. Otherwise you're right, 2.0 is 500MB/sec each direction.

aurgathor
join:2002-12-01
Lynnwood, WA

aurgathor to koitsu

Member

to koitsu
While I don't disagree that for some über geeks the two SATA III ports on the mobo may not be sufficient, I will stick to my original assertion that SATA III add-in cards in general make the most sense to those who have zero SATA III ports to begin with.

Previously, there used to be several other makers of PATA and SATA chips (i.e. Promise), but they don't seem to have many new consumer level products any more.

I'd be willing to pay over $100 for a good SATA III HA, but there are either the cheap, one chip solutions from several makers, or the much more expensive (often SAS) adapters starting around $400 - $500.

Yes, it would be nice if Intel would be selling separate SATA chips, but I'm afraid that we won't see those in the foreseeable future.

Having 2 chips with only 4 ports for an x4 bus is an interesting idea, though the 'right' solution would be to have a chip that's designed with 4 lanes.