<html><head><style type='text/css'>p { margin: 0; }</style></head><body><div style='font-family: arial,helvetica,sans-serif; font-size: 10pt; color: #000000'>David Shue will present his preFPO on Thursday April 25 at 9AM in Room 402.<br>The members of his committee are: Michael Freedman, advisor; Jennifer <br>Rexford and Anees Shaikh (IBM Research), readers; Vivek Pai and Margaret<br>Martonosi, nonreaders. Everyone is invited to attend his talk. His abstract <br>follows below.<br>--------------------------------------------<div style="color:#000;font-weight:normal;font-style:normal;text-decoration:none;font-family:Helvetica,Arial,sans-serif;font-size:12pt;"><div><font face="arial,helvetica,sans-serif" size="2">Title: Performance Isolation and Fairness for Multi-Tenant Cloud Storage</font></div><div><font face="arial,helvetica,sans-serif" size="2"><br></font></div><div><font face="arial,helvetica,sans-serif" size="2">Abstract:</font></div><div><font face="arial,helvetica,sans-serif" size="2"><br></font></div><div><font face="arial,helvetica,sans-serif" size="2">Shared storage services enjoy wide adoption in commercial clouds. But most systems today provide weak <br>performance isolation and fairness between tenants, if at all. Misbehaving or high-demand tenants can overload <br>the shared service and disrupt other well-behaved tenants, leading to unpredictable performance and violating SLAs.</font></div><font face="arial,helvetica,sans-serif" size="2"><br>In this thesis, we present Pisces, a system for achieving datacenter-wide per-tenant performance isolation and <br>fairness in shared key-value storage. Today’s approaches for multi-tenant resource allocation are based either <br>on per-VM allocations or hard rate limits that assume uniform workloads to achieve high utilization. Pisces provides <br>per-tenant weighted fair shares (or minimal rates) of the aggregate resources over the entire service, even when <br>different tenants’ partitions are co-located and when demand for different partitions is skewed, time-varying, or <br>bottlenecked by different server resources. Our key insight was to decompose the system-wide fair sharing problem <br>into a combination of four complementary mechanisms—partition placement, weight allocation, replica selection, <br>and weighted fair queuing—that operate on different time-scales and combine to achieve per-tenant max-min fairness. </font><div><font face="arial,helvetica,sans-serif" size="2"><br></font></div><div><font face="arial,helvetica,sans-serif" size="2">While Pisces provides fairness for key-value storage systems that leverage asynchronous writes for performance, <br>many storage systems require stronger durability guarantees and often provide rich data models that can support <br>arbitrary computation in the form of UDF's or map/reduce functionality. Achieving predictable performance in these <br>systems requires fine-grained per-tenant resource allocation over multiple resources (e.g network, disk, and CPU). <br>In the second part of this thesis, we generalize and extend the PIsces per-node scheduling model with Libra, a <br>multi-resource allocation library for building predictable shared services.</font></div><div><div><div><div><font size="4"><br></font></div></div></div></div></div><br></div></body></html>