Title A Cost-Effective, High-Bandwidth Storage Architecture Author Gibson, Nagle, Amiri, Butler, Change, Gobioff, Hardin, Riedel, Rochberg, Zelenka. Location ASPLOS Date 1998 Type of paper: new idea: using new technology (smarter hardware, disks) to take centralized servers (a bottleneck) out of the networked storage equation. Terminology: - object: file without a name, just an ID number. Disks essentially store object/offsets in a flat directory space. - SAN: Storage/System Area Network. Uses Fibrechannel which is subtlely different from TCP/IP. Fibrechannel has guaranteed delivery. - Capability: Clients contact servers to ask for permission to do something, and the server hand them a signed key (a capability) which the NASD verifies (because it knows the servers public key). - NAS: network attached storage, like an Network Appliance's NFS box. A NAS might speak NFS/AFS/CIFS and might keep a SAN behind it. Goals: Network Attached Secure Disk's (NASD) goals: 1) direct transfer to clients 2) secure interfaces via cryptographic support 3) asynchronous non-critical path oversight 4) variably-sized data object NASD separates management and filesystem semantics from store-and-forward copying. Figure 2 gives a fairly good synopsis of where how their architecture works as opposed to other forms. Using an object interface in storage rather than a fixed-block interface moves data layout management to the disk. The common case operation is read/write. Here we only need to go to the server for directory operations. Can putting the cryptography in silicon make it fast enough to decrypt on each block access? Problems: Security disabled in their experiments because their chips do not support it. Their AFS performance must suck because they say they built it but don't show it in their experiements. Intersting sources: Active Disk papers: Acharya98, Riedel98. Pannasus is a startup which is building NASDs. A Cost-Effective High-Bandwidth Storage Architecture The paper describes the Network-Attached Secure Disk storage architecture, which the authors believe provides scalable storage bandwith without the cost of servers used primarily for transferring data from peripheral networks to client networks. NASD separates management and filesystem semantics from store-and forward copying. It evolves the interface for commodity storage devices it eliminates the server resources required solely for data management. NASD has four main properties; it is a storage in which data is transferred between drive and client without indirection or store-and-forward through a file server machine. It has asynchronous oversight - the ability of the client to perform most operations without synchronous appeal to the file manager. NASD uses cryptographic techniques to defend against potential attacks from adversaries. This is particularly important since the storage is attached to the network and the drives are open to direct attack. The last important property of NASD is the object-based interface - drives export variable length objects instead of fixed-size blocks. This allows drives direct knowledge of the relationships between disk blocks and minimizes security overhead. NASD interface provides soft partitions, control objects and per-object attributes for data management. Objects with well-known names and structures allow configuration and bootstrap of drives and partitions. There is also an uninterpreted block of attribute space available to the file manager to record its own long-term, per-object state. NASD security is based on cryptographic capabilities. Capabilities are protected by a small number of keys organized into a four-level hierarchy. Clients obtain capabilites from a file manager using a secure and private protocol. Everytime when a request for an operation is issued by a client a driver verifies that request by checking the client's capabilites. The performance tests are not very optimistic. First, it turns out that the RPC mechanism is too costly so scalability suffers from that. The tests that estimate the computational power needed in a drive micro-controller show that NASD control is not too expensive but that workstation-class implementations of communications are. The authors show also that simple modifications to existing filesystems (NFS, AFS) allow NASD devices to be used without performance loss. The authors also build a small parallel distributed filesystem designed for NASD in order to show that NASD provides scalable bandwith. I need much more information about the internal operation of NASD in order to believe that it is a good idea. I would like to know how exactly the file manager works, how often it is queried, how much management power it has over the objects on disk. Also, I want to know more about the cryptographic techniques used and how much overhead they have. I believe most of this information is provided in other papers related to the same project. This paper gave me some general idea of the design of NASD but did not provide sufficient information that would let me evaluate it.