Alfred Perlstein

Projects:

• Scaling work:

  Tuned MySQL databases: analyzed slow queries, added indexes as needed to speed up queries, suggested rewriting application logic to make use of query cache, put mysql temp space on ramdisk to speed up non-memory-cache-able joins. Performed major MySQL data migrations to add required indexes and rebuild major tables.

  Tuned Apache and java instances to make use of more threads to reduce the per-process overhead for our internal caches.

• General systems work:

  Deployed an OpenStack cluster for use by team and by application team.

  Built out our Apache configuration system so that all configurations can be kept in a single repository and deployed to any host.

  Misc other stuff: setting up git repositories, deploying software, building packages, keeping people happy.

• Devops:

  Worked with application developers to learn enough about their software to be able to submit code changes/patches to fix issues across teams.

Projects:

• Scaling work:

  Added a reader/writer lock over the JUNOS networking stack to allow multiple networking flows to run in parallel. Achieved 5x speedup when processor was configured with 8 cores. Required back porting of FreeBSD 7.x rwlocks to FreeBSD 6 including turnstiles and sx locks and rwlocks. Required adding writer priority to the FreeBSD rwlocks.

  User space critical sections to prevent latency when contending against spin locks on cores running real-time processes.

  Flow affinity. Network processor pins flows to particular CPUs. By exporting this information to user space applications can exhibit much better cache and locking effects by assigning flows to threads already bound to the same CPU.

• FreeBSD liaison:

  Take non-core intellectual property modifications to FreeBSD and add them to FreeBSD to ease further upgrades of our code base.

  Recruited several strong FreeBSD developers into Juniper.

• Subversion deployment:

  Setup CVS to subversion gateway for team members to have branches. System tracked our CVS top of tree by committing to Subversion.

Projects:

• Match server:

  Created an linearly scalable system to replace a failure prone and non-scalable solution. Using two servers (but scalable to the hundreds) I fit a large subset of the users into physical ram using a custom R-star tree to locate users that are adjacent. Additionally a distributed cache (modeled after the memcached project) was modified to allow for distributed computation of match results.

  This system replaced a MySQL backed system and was able to perform a better job using half the hardware of the old system at a fraction of the complexity and maintenance of the old system.

  Additionally this system allowed for "fast path" access for several important data points for users.

• Image server migration and reliability:

  Our image servers were located at a separate hosting facility. Reliability problems at the hosting facility prompted me to move the machines to our local co-location facility. To maximize reliability, once the servers were made local, they were put behind a load balancer and made 100% redundant. Prior to this work, if an image server went down we lost images on the site until a manual switch over was done.

Projects:

• Autofs implementation: Completed a filesystem that supports dispatch of kernel filesystem operations to user land to facilitate dynamic browsing. This facilitates user land direction of the kernel policy presented to VFS consumers.
• Working on bug fixes and SMP support. Interests include NFS performance, software RAID, Symmetric Multiprocessing (SMP), the Linux emulator, kernel vm/buffer-cache, network subsystems and threads libraries.
• Assisted in the design of the multiprocessor safe (MP safe) network buffer allocator, currently working on making the socket subsystem MP safe.
• Made file allocation subsystem MP safe, collaborated with other FreeBSD developers (Seigo Tanimura in particular) to accomplish this.
• Ported BSD/os client side NFS locking implementation to FreeBSD. Included both user space and kernel subsystems.
• Added support for PCI based wireless cards based on the Prism/Prism2 chipsets.
• Modularized several large subsystems including the System V IPC subsystem and the POSIX aio subsystem by converting them to loadable kernel modules.
• PXE step-by-step guide to network booting FreeBSD systems for install or lab deployment. Presented work at Google (http://www.google.com) and Jarna (http://www.jarna.com/) . Project is available at http://people.freebsd.org/~alfred/pxe .
• Integrated and improved Yahoo! ( http://www.yahoo.com ) 'accept filter' kernel technology into FreeBSD.
• Ported pthread cancel functionality from OpenBSD

Projects:

• Session Server. A distributed database that stores key/value pairs in a highly replicated manner. Data is stored on multiple backends and the information needed to retrieve said data is made available to clients. Multiple machine failures do not impact the availability of data. Nodes participating in data storage are automatically detected via broadcast. Project included custom thread safe RPC server dispatch stubs and quorum logic.
• DNS Cache. A local cache for DNS requests designed to reduce the impact of latency and down time of DNS servers.
• Companion servers. A system for providing web ranks for users of the Yahoo toolbar. Optimized data build process from five days down to ten hours. Optimized lookup by storing data in binary format, tuning database cache size, splitting databases into multiple buckets. Tuned database build turaround from one week to ten hours.
• Invariants class. A C++ class developed in order to track down bugs in code. Utilized inheritance to provide all derived classes with checkpoints in order to ensure object coherence. Provided mechanisms for identifying corruption and memory leaks at runtime with negligible performance impact.
• Apache data bridge module. Created an apache module that made passing data between our core components and third party developer code simple. Greatly reduced the complexity required of third party developers interfacing with our code.

Projects:

• Main role assigned to enhance the experience of mobile users of NFS. This involved debugging several aspects of the NFS component mostly related to networking and process management.
• Provided implementation of autofs for OS X.
• Provided general filesystem notification mechanism for signaling filesystem alerts such as loss of connection to server, low disk space, filesystem mount and unmount. OS X is now the only UNIX based system to provide an escape mechanism for a broken NFS mount.
• Debugging, fixing and providing peer review to address issues. (example: forceful unmount of NFS filesystems.)

Projects:

• Developed TCP-splicing implementation. ftp://ftp.monarch.cs.cmu.edu/pub/dmaltz/msocks-infocom98.ps.gz ftp://ftp.monarch.cs.cmu.edu/pub/dmaltz/splice-perf-tr.ps

• Detailed code review facilitating the rewrite of high our premier product, a high performance web caching load balancer.
• Implemented non-blocking Sun RPC plug in to facilitate message passing API to reduce lock contention by assigning ownership of shared structures to a single process. This task was undertaken on my own in order to allow the product to be shipped without requiring interruption of service (reboot) for configuration changes.
• Developer assistance with workstation setup, using CVS, development software assistance (setting up X, vi/vim, emacs), understanding standard library calls and kernel facilities.
• Analyzing other projects being developed, finding and suggesting fixes for problems in code. Fixed memory corruption and leaks as well as file leaks in particular.

Projects:

• Lead designer of distributed high performance web tracking system. Project utilized C, Perl, SQL, PHP and FreeBSD.
• System administration, on call 24/7. Systems include high traffic web servers, DNS (bind 8), and NFS on platforms; FreeBSD (2.2.x/3.x/4.0), Sun Solaris (2.6/7) and Redhat Linux.
• Maintaining systems and source code via CVS.
• Designed an interpreter used to increase page rankings for web pages.
• Maintained systems consisting of databases (My-SQL, Postgresql), Perl, php and sh scripts. Responsible for backups, system security and bandwidth monitoring.
• Modified Postgresql client libraries to provide nonblocking support for C client code. (integrated in mainstream sources)
• Added support to pre-load data into the FreeBSD kernel to support zero copy data transport. Achieved 2500 connections per-second on a single 400mhz PII.
• Building x86 and Sun Ultra-sparc systems using FreeBSD and Solaris for internal and customer use.
• Tuning FreeBSD systems to support large Postgresql and Apache installs.