technical info: alternative controllers for the BackBlaze Pod design

[The O.G. X]

I don't know what alternative controllers others have tried, but we've been experimenting a bit and wanted to share with the community our findings:

1) HighPoint Rocket 640L controller (this is different from the RocketRAID). This controller is based on the marvell 88SE9230 chipset and has some rudimentary support for port multipliers like the Sil one that the backblaze design uses. However, for some strange reason, we could only it get it to recognize a maximum of 4 drives. Might be related to firmware, etc.... but it does seem to "sort of" work with Sil based port multipliers. The reason we looked into this is because of the 2x lane PCI-E 2.0  and SATA-3 support, which we were hoping to increase the pipe between the drives in the backplane to the motherboard. We were hoping this would work, it's a cheap card, about $60 each. Perhaps with a firmware upgrade it could support 4x5=20 drives?

2) HighPoint RocketRAID 640L controller. This one is similar to the above controller, but with built-in RAID support *and* official port multiplier support from the manufacturer. They claim, with 1x5 port multipliers, the single card can support 20 hard drives with the 4 SATA connectors it has. So, sounds promising and we're receiving these today to test. This too is a 2x lane PCI-E 2.0 with SATA 6Gbps support. I'll post with our results but we're hoping this will work. If this work, and provide better bandwidth, it's only $100 per controller so still a reasonable alternative, if it works.

What other SATA controllers have people tried with the BackBlaze design?

[maximilie...@gmail.com]

Hi The O.G. X,

Any new feedback ?
Thanks in advance,
Max.

[Tim Small]

 based on the marvell 88SE9230 chipset

I've had poor experience with the Marvell based chips (see my recent post which really should have been in this thread - sorry!).  Here are a couple of threads related to Marvell + port multiplier support:

http://marc.info/?t=130747326400001&r=1&w=2

http://marc.info/?t=130907214900002&r=1&w=2

I've also seen controller lockups on the Marvell AHCI controllers unless NCQ is disabled.

SATA Port multipliers seem to provide limited bandwidth and are a little fiddly to support in hardware and software, so I'm wondering if a solution based on PCIe switches wouldn't be better?  You can then use the fast/wide PCIe slots that many motherboards have at least a couple of, together with commodity SATA controllers.

The sil3132 is pretty good, but is only PCIe 1.x and coming to end-of-life.

The sil3124 is a PCI-X design now fronted by PCIe to parallel PCI bridge chips, and are either very expensive (x4 designs) or slow, and seem like a weird way to do things.

I've found the Marvell chips unreliable (perhaps newer ones are better, but I just didn't like their general attitude much - e.g. no errata released except under NDA!).

Similar with LSI (buggy and bad attitude).

The ASMedia 1061 seems promising, but I haven't any experience of it with SATA port multipliers (to fit in to 1.0 -> 3.0 pods), very cheap with good performance - but limited to 2 ports (PCIe 2.0 single lane), so of limited use unless behind a PCIe bridge

...so a PCI switch based design would need a custom PCB, or maybe at a push something like a  bunch of these: http://eshop.sintech.cn/pcie-express-3port-1x-multiplier-riser-cable-diy-bitcoin-miner-p-1011.html (n.b. this isn't USB3.0, the design just uses a USB3.0 cable as a physical transport for the PCIe signal) - at least as a proof of concept...  Approx cost $100 for 6 ports (plus some mounting hassles).

Backblaze have gone for the Rocket 750 of course - this uses a Marvell 88SE9485 (8 port SATA/SAS PCIe 2.0 x8), together with (I assume) eight 5 port Marvell port multipliers - giving 120 MB/s per SATA port.  Pricey tho', and Marvell have given me so much hassle in the past that I'm reluctant to go with them again...

"SATA Express" (essentially hard drives which speak PCIe directly instead of SATA) may make things more interesting, if the uptake is quick enough...