CS E6998-10. Advanced Topics in Network Storage Systems Midterm Exam, March 9, 2004 Name: ID#: 1. a. Briefly describe the migrations taking place internally between the two storage hierarchies in an HP AutoRAID system. Under what conditions is an HP AutoRAID system expected to be thrashing? (Hint: Think about application workload characteristics) Migration from RAID5 and RAID1 takes place when data in RAID5 is overwritten. Migration from RAID1 to RAID5 is based on a least-recently written algorithm. If the active-write working set exceeds the size of mirrored storage for long periods of time, it is possible to drive the AutoRAID array into thrashing mode, in which case each update causes the target RB to be promoted up to the mirrored storage and a second RB demoted to RAID5. b. Assuming that a storage system administrator cannot interfere with applications already running and using the HP AutoRAID but has the ability to add/remove storage resources to it at any time, which action would you recommend to remedy the thrashing condition and why? Add disks. 2. Briefly descibe the most important high-availability features of the IBM Enterprise Storage Server (ESS). Each ESS unit has two SMP nodes. Each node "owns" half of all disks but can potentially access all disks and can take-over in case of failure. Device Adapters are paired. Configuring RAID arrays over the back-end disk groups provides high-availability at the disk level. 3. Briefly describe the most significant point of departure of the redesigned AFS from the prototype regarding cache management. The redesigned AFS uses callbacks. No need to validate at each open. 2. a. Describe the chained declustered layout used in Petal. What are the benefits and drawbacks of this layout over a standard mirrored layout such as RAID-1? See Petal paper Section 2.4 and Figure 5. Benefits: Better load-balancing after disk failure. Drawback: Less reliable. b. Petal supports snapshots of virtual disks. This capability can be used to create a snapshot a Frangipani file system for backup purposes, without suspending or otherwise disturbing normal Frangipani operation. Describe how such a Frangipani file system can be restored from a backup snapshot at a later time. A Petal snapshot of a Frangipani file system is "crash-consistent". To restore the file system from such a snapshot it is necessary to run the Frangipani logs. c. Your employer asks you to implement a variant of HA-NFS by replacing the original HA-NFS storage backend with Petal. - Describe how you would implement such a variant without inflicting major changes to HA-NFS When the HA-NFS volumes are stored in Petal instead of the standard dual-ported disks, both servers can still access each other's volumes by virtue of Petal's shared disk abstraction. So, they can still take-over in the case of a server failure, i.e., run the other server's logs, etc. Note that a SCSI channel between the server may still be necessary as an additional check for liveness (besides RPC null and ICMP ping tests). - How can you take advantage of Petal to achieve fault tolerance against types of disasters the original HA-NFS was vulnerable to? Geographic distribution possible with Petal enables HA-NFS to tolerate entire site failures, e.g., a power failure or a natural disaster that brings down an entire site in a single geographical location. 3. a. Which bottleneck of traditional network file systems is NASD designed to eliminate? Judging from experimental measurements in the Gibson paper, how would you characterize the scalability of a NASD system? (Hint: Think of a common characteristic in the NASD scalability measurements, ie. graphs where the number of clients is varied in the x-axis). NASD eliminates the file server bottleneck. Scalability is linear. b. Explain why a NASD-based storage system may be more cost-effective than a standard network file system such as NFS? You no longer need to pay the cost of a high-end file server. c. How does NASD ensure that a client access to a NASD drive is authorized? In your description, make sure you mention all parties involved in this process, from the time the authorization is issued to the time the authorization is verified. The NASD file manager issues a capability and hands it to client at open time. The NASD drive can verify that capability is valid. The NASD File manager and drives share a secret key. 4. a. What are the characteristics of the SPECsfs performance curve? Draw a hypothetical such curve. Base, slope, saturation point. See Martin paper for details. b. How is an NFS server modeled analytically in the Martin paper? Draw a figure describing the model. See Martin paper for details (Figure 3). c. Which network parameter has the highest impact on NFS performance? Which part of the SPECsfs performace curve is most affected by changes to this network parameter? Overhead. Mostly affects saturation point (i.e., peak ops/sec). 5. a. How does the Direct Access File System (DAFS) avoid data movement costs in network file access? Uses remote direct memory access (RDMA) to move data directly between the network and application buffers. b. How can applications benefit from reduced data movement costs when using DAFS? Reduced data movement costs imply lower overhead, which means more CPU cycles left to be used for application processing. c. Besides the reduction of data movement in network file access, what is the other defining characteristic of the DAFS protocol and how can applications benefit from it? User-level file client can be used for full control over I/O policies. Better caching, prefetching policies. Better portability. 6. a. What are the advantages of using the iSCSI protocol to access a SCSI target device over using FibreChannel? iSCSI can be used over existing network infrastructure. Since it is layered over TCP/IP, it can easily be used over longer geographical distances. b. In the iSCSI protocol, why is it desirable to use multiple TCP connections, preferably over multiple physical interconnections between two hosts? Multiple TCP connections can better utilize the full link bandwidth. Multiple physical interconnections can be used for failover in case of link failure.