Glauber Costa on OSv as NoSQL platform
2 likes1,634 views
Cloudius Systems is a startup focused on creating an OS designed for public and private clouds, specifically a lightweight 'library OS' built on C++11, which aims to simplify application deployment directly onto hypervisors. The performance improvements include reduced system call costs, fast networking, and faster boot times, with tools like Capstan for application management. Though OSv currently supports Java applications with some NoSQL databases, ongoing developments aim to enhance performance and compatibility with other systems.
![What do people think?
Roman Shaposhnik, Bigtop/Hadoop
"Mark my words, GridGain [...] and OSv [...] are going to
be excitingly disruptive in the next few years. […] And,
by the way, if, after reading this blog, you are not
dropping everything and porting your cloud application
to OSv, I don’t know what’s wrong with you."
http://blog.gopivotal.com/pivotal/features/pivotal-debuts-at-apachecon-north-america-2014-thanks-for-having-us](https://image.slidesharecdn.com/nosqlmatters-140430134208-phpapp02/85/Glauber-Costa-on-OSv-as-NoSQL-platform-7-320.jpg)
Glauber Costa on OSv as NoSQL platform
- 1. Glauber Costa, Lead Engineer glommer@cloudius-systems.com
- 2. Who are we ● Small startup founded by Avi Kivity (Creator of KVM) ● Around 15 people, including some Linux veterans ○ 10 countries ● Headquarters in Israel ● Our mission is to build the default OS for public and private Clouds.
- 3. Typical Cloud Stack Application Application Server Runtime (JVM) Operating System Hypervisor Hardware Protection and abstraction
- 4. “Library OS” Application Application Server Runtime (JVM) + OSv Hypervisor Hardware
- 5. OSv ● LibraryOS written from scratch in C++11 ○ Actually network stack came initially from FreeBSD, heavily modified ○ ZFS filesystem from Open Solaris ● BSD Licensed. ● Implements most of the (Linux) POSIX API ● Better if we have a runtime (Java at the moment) ● It supports only one application. ● KVM, Xen, VirtualBox, VMWare ● Image size as low as ~15 Mb ● Beta expected in a couple of months.
- 6. OSv and NoSQL ● Think of us as "No-OS" ● You can distribute a copy of your application that will run directly on a hypervisor/cloud ● Cassandra ○ Due to its Java focus ● in memory redis ○ Due to its pervasiveness ● memcached ○ Due to its simplicity ● mongodb ○ Because I wanted to.
- 7. What do people think? Roman Shaposhnik, Bigtop/Hadoop "Mark my words, GridGain [...] and OSv [...] are going to be excitingly disruptive in the next few years. […] And, by the way, if, after reading this blog, you are not dropping everything and porting your cloud application to OSv, I don’t know what’s wrong with you." http://blog.gopivotal.com/pivotal/features/pivotal-debuts-at-apachecon-north-america-2014-thanks-for-having-us
- 8. Performance ● System calls are free ○ context switches are really cheap ( x 4 Fedora* Linux) ● Network performance significantly faster ○ around 20 % with netperf over Fedora Linux ○ more than 50 % for some UDP workloads ● SpecJVM between 3 and 5 % faster. ○ Not a lot actually expected ● Boots < 1 second. * Fedora 19, roughly 6 - 8 months old Linux Kernel
- 9. Real savings: sysadmins ● No command line, ○ except for compatibility ● no graphical interface either. ● REST API for full automation. ● Almost zero configuration ○ Compare with almost 200 in a very stripped down Linux ○ or even registry.
- 10. Capstan ● Capstan is a tool for building and running your application on OSv (under a hypervisor) ● Docker-like command line interface ○ capstan run cloudius/cassandra ● Multi-platform ○ Linux, OS X, Windows ○ QEMU/KVM, VirtualBox, VMware ○ Google Compute Engine, Amazon EC2
- 11. mmap ● mmap is a system call for mapping files to memory ○ NoSQL databases rely on it for persistence, caching, I/O, and copy- on-write ● Page cache behavior is important for database performance ○ Recent Linux & PostgreSQL fsync() problems ● ZFS + mmap = :-( ○ ZFS ARC cache is not integrated with page cache. ○ We are trying to fix that.
- 12. memcached ● unmodified: ● modified: ○ that showcases the real performance opportunity behind OSv. That's already ~20 % more at http://osv.io/benchmarks/
- 13. Cassandra ● A bug in our mmap implementation creates some instability ● Performance is on pair with Linux ○ But room for improvement is very big. ● There is, however, less need to worry about JVM parameters due to JVM Ballooning and friends
- 14. MongoDB, redis and more ● recently started to meddle with them ○ No trusted numbers for them yet - follow us for news! ● Mongo has a problem with mmap as well as Cassandra ○ Should be ready soon ● Redis is way too dependent on COW/fork for persistence ● Mongo and Cassandra are quite dependent on mmap ○ There is still work to do with that, but over time using our own APIs can help there as well
- 16. Virtualization Oriented ● No spinlocks in the whole kernel ○ Because of lockholder preemption ● No complicated hardware model ● Avoid things that are traditionally more expensive in HV ○ Like IPIs and timer setting ● We also feature a fair scheduler, support multiple page sizes transparently and avoid page metadata overhead.