4 ms·

Getting Physical - Seagate Kinetic Drive Hardware

15 points by jarnold 13y ago

ambiate 13y ago
RAID over IP via broadcast? From a security standpoint, it would seem there are many hurdles to jump. Not too long ago, I talked to Dickinson about what you guys do. I never considered the amount of harddrives necessary to facilitate Swiftstack's goals. Of course the software challenges are interesting, but the amount of hardware and maintaining risk/DR/costs/etc. Oh man. Hats off to you guys.
- notmyname 13y ago
  It's not RAID over IP. Each drive itself is a network endpoint. On the practical side, this means that the drive is "connected" to one or more servers (and can be re-homed on the fly). The storage system itself is responsible for coordinating the communication across the cluster of Kinetic drives. In our case, the coordinating storage system is OpenStack Swift. On the humorous side, this means that technobabble like "Can you ping the boot record?" or "What does a traceroute to the directory show?" actually sorta make sense now.
  - ambiate 13y ago
    RAID over IP was a joke. I remember reading this story about someone flooding routers to turn them into a broadcasts to view the flow of packets over the network. I imagined the same thing, then said, hey with HDDs and broadcast, you could simulate RAID. I made a long stretch on that one.
    - mbreese 13y ago
      That's probably not too far off. If you wanted to do mirror a write, you could imagine broadcasting that same packet to a set of devices. The devices would have to figure out whether or not they needed to store the packet. Otherwise, you'd be sending N packets to the same switch for N-way mirroring. In a typical hardware RAID, the sector write is sent to the RAID controller once. It's probably not too long of a stretch.
      - toomuchtodo 13y ago
        Writing with multicast and using tcp only for the control plane? Interesting concept.
  - wires 13y ago
    lol @ technobabble
mrbill 13y ago
Same connector as SAS, but a different pinout? Does this not scream "bad idea" to anyone else?
- jarnold 13y ago
  This is a good thing for a couple of reasons. First it allows equipment manufacturers to build chassis with the same drive trays that are already in production. Second, it that from a drive assembly standpoint, the same build and testing equipment can be used. This should speed development and manufacturing quality. Is the concern more from an ergonomic perspective, that someone in a data center may plug in a drive into the wrong system?
  - mbreese 13y ago
    If the testing equipment is expecting a SAS port, then wouldn't the dual ethernet ports pose a problem? Or, is it that they are using SAS power pins to do the ethernet transfer leaving the other pins untouched? (I don't know my SAS pin-outs very well). The main concern would be that, yes, someone could plug the drive into the wrong system.
  - deelowe 13y ago
    I'd imagine that would be a pretty big concern. This is a regression for serviceability. Connectors should be keyed specific to the function for a variety of reasons. I'm sure seagate has their reasons, but this seems like a bad idea to me.
- XorNot 13y ago
  That was my first thought. Wonder what the failure mode is if you mismatch things?
  - jarnold 13y ago
    Well given that the device will start asking for a DHCP address when it is powered on, it will probably not do anything.

15 points by jarnold 13y ago

RAID over IP via broadcast? From a security standpoint, it would seem there are many hurdles to jump. Not too long ago, I talked to Dickinson about what you guys do. I never considered the amount of harddrives necessary to facilitate Swiftstack's goals. Of course the software challenges are interesting, but the amount of hardware and maintaining risk/DR/costs/etc. Oh man. Hats off to you guys.

notmyname 13y ago
It's not RAID over IP. Each drive itself is a network endpoint. On the practical side, this means that the drive is "connected" to one or more servers (and can be re-homed on the fly). The storage system itself is responsible for coordinating the communication across the cluster of Kinetic drives. In our case, the coordinating storage system is OpenStack Swift. On the humorous side, this means that technobabble like "Can you ping the boot record?" or "What does a traceroute to the directory show?" actually sorta make sense now.
- ambiate 13y ago
  RAID over IP was a joke. I remember reading this story about someone flooding routers to turn them into a broadcasts to view the flow of packets over the network. I imagined the same thing, then said, hey with HDDs and broadcast, you could simulate RAID. I made a long stretch on that one.
  - mbreese 13y ago
    That's probably not too far off. If you wanted to do mirror a write, you could imagine broadcasting that same packet to a set of devices. The devices would have to figure out whether or not they needed to store the packet. Otherwise, you'd be sending N packets to the same switch for N-way mirroring. In a typical hardware RAID, the sector write is sent to the RAID controller once. It's probably not too long of a stretch.
    - toomuchtodo 13y ago
      Writing with multicast and using tcp only for the control plane? Interesting concept.
- wires 13y ago
  lol @ technobabble

mrbill 13y ago

Same connector as SAS, but a different pinout? Does this not scream "bad idea" to anyone else?

jarnold 13y ago
This is a good thing for a couple of reasons. First it allows equipment manufacturers to build chassis with the same drive trays that are already in production. Second, it that from a drive assembly standpoint, the same build and testing equipment can be used. This should speed development and manufacturing quality. Is the concern more from an ergonomic perspective, that someone in a data center may plug in a drive into the wrong system?
- mbreese 13y ago
  If the testing equipment is expecting a SAS port, then wouldn't the dual ethernet ports pose a problem? Or, is it that they are using SAS power pins to do the ethernet transfer leaving the other pins untouched? (I don't know my SAS pin-outs very well). The main concern would be that, yes, someone could plug the drive into the wrong system.
- deelowe 13y ago
  I'd imagine that would be a pretty big concern. This is a regression for serviceability. Connectors should be keyed specific to the function for a variety of reasons. I'm sure seagate has their reasons, but this seems like a bad idea to me.
XorNot 13y ago
That was my first thought. Wonder what the failure mode is if you mismatch things?
- jarnold 13y ago
  Well given that the device will start asking for a DHCP address when it is powered on, it will probably not do anything.