https://homeostasis.scs.carleton.ca/wiki/index.php?action=history&feed=atom&title=DistOS_2021F_2021-11-09 2026-05-12T18:52:28Z Revision history for this page on the wiki MediaWiki 1.42.1 https://homeostasis.scs.carleton.ca/wiki/index.php?title=DistOS_2021F_2021-11-09&diff=23449&oldid=prev 2021-11-11T21:48:02Z

<p>Created page with &quot;==Notes== &lt;pre&gt; Lecture 15 ---------- - experiences - proposal - midterm update - participation Spanner - a big, distributed (semi-)relational database - very consis...&quot;</p> <p><b>New page</b></p><div>==Notes==<br /> <br /> &lt;pre&gt;<br /> Lecture 15<br /> ----------<br /> - experiences<br /> - proposal<br /> - midterm update<br /> - participation<br /> <br /> Spanner<br /> - a big, distributed (semi-)relational database<br /> - very consistent<br /> - supported SQL<br /> - all of the query parts<br /> - management, maybe not so much?<br /> - big deal, because of usability<br /> - developers know SQL<br /> - want transactions, helpful for consistency<br /> across tables<br /> <br /> in distributed systems, we&#039;re always making tradeoffs<br /> between functionality, scalability, and complexity<br /> - normally we just think about functionality vs scalability<br /> (SQL vs NoSQL)<br /> - but add complexity and you can get functionality &amp; scalability at the same time<br /> <br /> <br /> Spanner is proprietary to Google, others have<br /> made their own versions (CockroachDB)<br /> <br /> <br /> Tradeoff also shows up in Tensorflow<br /> - for &quot;machine learning&quot;<br /> - what is it really for?<br /> - working with n-dimensional arrays (i.e. tensors)<br /> - and we can do neural networks if we can do fast<br /> tensor processing<br /> <br /> Is this just the same thing as MapReduce?<br /> - what&#039;s different?<br /> - not embarassingly parallel!<br /> - have to communicate between tasks as they run,<br /> not just at the end (i.e., during reduce)<br /> <br /> Modern machine learning is based on large, mutable models<br /> - MANY parameters (weights in the neural network)<br /> <br /> Basic idea of a neural network<br /> - input, hidden, and output nodes<br /> - input nodes are connected to layers of hidden nodes<br /> - hidden nodes are connected to output nodes<br /> - weights on connections between nodes determine<br /> how values are transformed as they propagate along connections between nodes<br /> <br /> So here, take an input tensor, transform it a bunch of times until you get an output tensor<br /> <br /> All that &quot;deep&quot; learning means is it is a neural network<br /> with many, many layers of hidden nodes<br /> <br /> The cool part about tensorflow is you don&#039;t care about the hardware<br /> - your data model can be efficiently mapped onto a wide<br /> variety of architectures<br /> - big change from past efforts in supercomputing<br /> &lt;/pre&gt;</div>

Soma