Outline • PAST goals Storage management and caching • PAST api in PAST • File storage overview • File and replica diversion Antony Rowstron and Peter Druschel • Replica management Presented to cs294-4 by Owen Cooper • Caching • Performance • Discussion PAST (non)goals Security Model • P2P global storage network • Pastry node ids are a hash of a public key – Use properties of existing p2p systems (Pastry) • Smartcard based security – Support for strong persistence – Provides keys • Via a core set of replicas – High availability – Quota management • Via local caching • Nodeid and fileid generation controlled – Scalable – Try to stop nodes from getting consecutive ids • Obtain high storage utilization via local cooperation – Or clients from overloading parts of the network – Secure • Design goals do not include • But node id and real world identity may not be – Replacing the file system linked – Updatable files • Data not encrypted – Directory or lookup service
PAST API’s File insertion • In PAST, files are immutable • Insert(name, c, k, file) • Fileid=Insert(filename,credentials, k, file) – Computes a storage certificate • Contains fileid, hash of content, k, salt – Insert k copies of the file into the network, or fail. – Deducts k*filesize from quota – Fileid a signed (filename, credentials, salt) – Routes file and storage certificate using pastry using – Successful if ack with receipts from k nodes fileid. • File=lookup(fileid) – Node verifies the integrity of the file, stores it, and asks – Return a copy of the file if it exists k-1 closest nodes to store the file. • Reclaim(fileid, cradentials) • K-1 nodes in leaf set (k-1 <= l) – Node returns ack with k signed storage receipts, or a – Reclaim accepted if requested by the owner nak. – Allows, but does not require, storage reclamation Lookup and Reclamation Diversion • Pastry ensures replica is found • A file or replica can be relocated • For a replica, to another close node – Since a lookup is routed to the closest nodeid – If one of the K closest is overloaded • Reclamation • For a file, to another set of nodes in the idspace – Client generates a reclaim certificate – If the nodes around a fileid are (possibly locally) congested – Sends it to the fileid via pastry • Why is this necessary? – Recipients verify the certificate & issue receipt – Differing storage capacity at nodes – Client reclaims quota – Differing file size for inserted files
Replica Diversion File Diversion • Node responsible for fileid asks k-1 neighbors to • Replica diversion is local store the file – Allows storage choice between nodes around • Neighbor (N) may divert a copy to a node in its fileid leaf set • File Diversion – Pointer to copy inserted at N – Triggered when an insert with a fileid fails – N issues storage certificate – Insert is tried a total of three times – N also inserts a pointer on the k+1th closest node – New fileid generated by changing the salt • No orphan if N fails • N remains responsible for pointer maintenance Storage Policy Replica maintenance • How does a node choose to accept or reject a • Node join/leave causes responsibility shift replica? – Pastry node failure detection will cause leaf set updates – Computes sizeof(file)/sizeof(free_space) • Past detects responsibility shifts this way • Newly responsible node must copy files – Compares to T pri or T div depending node’s role – T pri > T div – Make a copy immediately, OR • How is node chosen for replica diversion – pointer to old owner & copy lazily • Diverted replicas – Search leaf set for the node that • Has maximal free space – Target of diversion may move out of leaf set • Doesn’t already hold a diverted or primary replica • Node to store repica can be any one in leaf set • File diversion – Must exchange keepalive messages themselves – K copies cannot be located (via primary or diversion) – Should be relocated
Replica maintenance (2) Caching • Pastry’s locality based routing will tend to direct • Node failure may cause storage shortage requests to nearby copies – No node in leaf set can take over ownership • PAST also stores cached copies • Search space is widened – Along routing path between client and fileid – Ask most extreme nodes to locate storage – For insert and lookup operations – Cache maintained using GD-size algorithm • Increases search space to 2l nodes • Weight per file: 1/size(file) – If no storage space found, fail. • Eviction: – Pick file with minimum weight – Subtract weight of evicted file from all others Experiments: without diversion Experiments (2): with diversion • Experiments use – Large trace from web server – Files from local web server • The case for diversion with web trace – Without diversion: • 51.1% of insertions failed • 60.8% storage utilization • With diversion – Bigger leaf set size a plus
Experiments (3):varying T pri Experiments (4): Varying T div • Varying T div • Effects of varying T pri • T pri is constant • # files stored v.s. size of file File and Replica Diversion caching • 8 traces combined • Requests from clients in each trace are mapped to close PAST nodes
Recommend
More recommend
