Advances, Systems and Applications
Table 5 Summary of further analysis
Work | Data transfers and imbalance | Dynamic scheduling | Hybrid and Multicloud | Workflow support |
|---|---|---|---|---|
PANDEY et al., 2010 | Transfers are evaluated via workflow DAG and resource allocation; transfer imbalance is not addressed. | Only addresses fluctuations in the transfer costs. Other aspects such as performance fluctuations and reliability are not mentioned. | No explicit support or experiments. | Modeled as DAGs; richer characterizations are not supported. |
LIN; LU, 2011 | Transfer capacity of nodes in the same network are assumed to be uniform. Transfer imbalance is discarded. | Not addressed. | No explicit support or experiments | Supported; no details included. |
XU et al., 2009 | Transfers and data properties are not explicitly addressed. | Not addressed. | No explicit support or experiments. | Multiple workflows supported via common merging point; simple DAG modeling. |
WEISSMAN; GRIMSHAW, 1996 | Data locality is a scheduling constraint; worker must be assigned closer to data. | Two levels: local and global. Rescheduling is first handled on local level. Details are not provided. | Design for wide-area systems (pre-dates cloud computing). | No explicit support. |
CHEN; ZHANG, 2009 | Data communication and transfers are not explicitly addressed. | Not addressed. | No explicit support or experiments. | Simplified DAG model without edge costs. |
RODRIGUEZ; BUYYA, 2014 | Rigidly modeled; fixed costs for transfers and no cost for local I/O. | Not addressed. | Not addressed. | DAG with fixed transfer costs and computation costs based on FLOPS. |
FARD et al., 2012 | Transfers are considered but contention effects are not. Energy calculations ignore transfer times. | Not addressed. | No explicit support or experiments. | DAG with fixed transfer costs; not details on task costs. |
MALAWSKI et al., 2012 | Algorithm does not consider the size of input data; transfer time is part of computation. | Initial scheduling plus periodic adjusting depending on amount of idle resources. | No explicit support or experiments. | DAG with fixed transfer costs and computation costs with slight variability. |
SAKELLARIOU; ZHAO, 2004 | Linear variation to amount of input data size. | Immediately before execution of tasks and bound to a condition to minimize number of reschedules. | Not addressed; solution originally designed for grids. | DAG with computation and transfer costs modeled with linear variation w.r.t. amount of input. |
WANG; CHEN, 2012 | Not addressed. DAG does not specify transfer costs. | Not addressed. | No explicit support. | DAG with tasks and implicit costs. No transfer costs and no more complex characterization. |
POOLA et al., 2014a | Based on data size and one value for network bandwidth. | Not addressed. | No explicit support. | DAG with task cost based on number of instructions. |
BITTENCOURT; MADEIRA, 2011 | Based on data size and fixed network bandwidth values among nodes. | Two-step scheduling: static, then including public cloud to address deadline. | Initial scheduling step considers private resources; public resources are used if necessary. | DAG with compute cost based on number of instructions. |
VECCHIOLA et al., 2012 | Not specified. | Not addressed. | Public resources used if necessary. | Supported, but no details provided. |