1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
|
SERVICE/WORKER ARCHITECTURE
* pause/resume job
* cancel job
* cancel map
* slow node
* one data point more expensive than the others
* infinite loop for some values
* exceptions for some values
* worker not stateless: one calculation interferes with the next
* worker memory leak
* worker requires initial state
* worker state changes between map calls
* client requests logging on service, any worker, or all workers
* logging based on condition
* rerun value with logging when worker fails
* parallel random number generator, with seed control
* worker which needs workers
CLUSTER MANAGEMENT
* priority jobs
* start/stop workers when job starts
* add client to processing pool, with preference for client jobs
* add new node
* machine reboots: exchange, service, worker, monitor
* unreliable processor/memory on some nodes
* workers which use local disk but don't clean up after themselves
* service monitoring
* identify all processes for a given user
* identify idle machines
JOB MANAGER
* job manager machine reboots
* job manager upgrades
* progress thumbnail
* client detach/attach
* retrieve results
* notify user when job starts (message queue or email)
DEPLOYMENT
* egg basket and trusted users
* different users having conflicting dependencies
* worker implemented in java, IDL, Matlab, R, ...
* client implemented java, IDL, Matlab, R, ...
* code movement while developing worker
DATA MANAGEMENT
* big file, independent disks
* storing run results
SCALABILITY
* workstation, dedicated cluster, batch queue, NoW, EC2, teragrid
SECURITY (SNS, Teragrid, shared clusters)
* service/workers run as user
* exchange runs as user
* job manager behind firewall
* cluster behind firewall
|
