Skip to content

Latest commit

 

History

History
178 lines (145 loc) · 8.6 KB

INSTALL.md

File metadata and controls

178 lines (145 loc) · 8.6 KB
Raw

Installing Stenographer

If you'd prefer to read a shell script, you can take a look at install.sh :) Also, we do plan to eventually create a real debian package config, and once that's done we'll provide deb packages for easier installation.

This documentation provides our current method for installing stenographer on a machine, including the justifications for why we think that method is currently best-practice (or why it's not ;)

User/Group Setup

We set up a single system group stenographer, and a system user stenographer.

Group stenographer

The stenographer group is used to control access to locally stored packet data. Users are added to this group to allow them to query stenographer (via the stenoread command).

User stenographer

The stenographer user is used to run the stenographer and stenotype binaries. We use the stenographer user to protect the system from stenographer and stenotype... this user has no special privileges except the ability to read/write packet data and run the (setcap'd) stenotype binary. So if either is compromised, the system as a whole won't be. See the Defense In Depth section of DESIGN.md for more details.

Configuration Files

There are a number of files in the configs/ subdirectory, which may help in installation.

  • steno.conf: Discussed in more detail in the Configuration File section. A configuration file must exist at /etc/stenographer/config
  • upstart.conf: Upstart configuration file, can be copied into /etc/init/stenographer.conf to allow upstart to manage Stenographer
  • limits.conf: If you don't use upstart, you may need to move this to /etc/security/limits.d/stenographer.conf in order to allow Stenographer to create files at the size it needs, and to open the number of files it needs to

Needed Directories

There are a few directories Stenographer needs in order to run correctly:

  • /etc/stenographer root:root/0755: Stores configuration file
  • /etc/stenographer/certs stenographer:stenographer/0750: Stores certificates used to verify clients are allowed to access packet data. stenographer writes certificates for client and server to this directory, and read clients stenoread/stenocurl read these certs and use them to make requests.
  • Packet directories: These are chosen by the installer. See Configuration File section for more details.

Configuration File

The /etc/stenographer/config file tells Stenographer what packets to read, where to write them, how to serve them, etc. It also tells the clients where the Stenographer server is running and how to query it.

Here's an example config (note: it's JSON):

{
  "Threads": [
      { "PacketsDirectory": "/disk1/stenopkt", "IndexDirectory": "/disk3/stenoidx/disk1"}
    , { "PacketsDirectory": "/disk2/stenopkt", "IndexDirectory": "/disk3/stenoidx/disk2", "DiskFreePercentage": 25}
  ]
  , "StenotypePath": "/usr/local/bin/stenotype"
  , "Interface": "em1"
  , "Port": 1234
  , "Flags": []
  , "CertPath": "/etc/stenographer/certs"
}

Let's look at each part of this in detail:

  • StenotypePath: Where stenographer can find the stenotype binary, which it runs as a subprocess
  • Interface: Network interface to read packets from
  • Port: Port stenographer will bind to in order to serve stenoread requests.
  • CertPath: Where stenographer will write certificates for client verification, and where the clients will read certificates when issuing queries.

Threads

The Threads section is one of the most important. It tells stenotype, the packet capturing subprocess, a number of things: how many threads to read packets with, where to store those packets, and how to clean them up.

For each packet reading thread you'd like to run (IE: for each core you'd like to use), you must specify:

  • PacketsDirectory: Where to write packet files. We recommend mounting a separate disk for each thread... we've found that at least for spinning disks, a single core can easily fill a disk's entire write throughput with room to spare.
  • IndexDirectory: Where to write index files. We've had good luck with using a single separate disk to write all index files, writing each thread's index to a separate subdirectory. This directory gets FAR fewer writes, and they're FAR smaller. We've found that even with up to 8 threads, the all 8 index directories take up less than 20% of the space of a single thread's packets.
  • DiskFreePercentage: The amount of space to keep free in the packets directory. stenographer will delete files in this thread's packets directory when free disk space decreases below this percentage. Note that we don't currently do any automated cleanup of the index directory. When a packet file is cleaned up, its index file is cleaned up, and because index files take so little space, we haven't ever needed to clean them up directly. Note that DiskFreePercentage is optional... it defaults to 10%.
  • MaxDirectoryFiles: The maximum number of packet/index files to create before cleaning old ones up. Defaults to 30K files, to avoid issues with ext3's 32K file-per-directory maximums. For ext4 you should be able to go higher without issue. Note that since we create at least one file every minute, this defaults to a maximum limit of 8 1/3 days before we drop old packets.

Flags

The Flags section allows you to specify flags to pass to the stenotype binary. Here are some flags which may prove particularly useful:

  • -v: Add verbosity to logging. Logs by default are written to syslog, and are relatively quiet. Adding one -v will have stenotype write per-thread capture statistics every minute or 100MB of packets, whichever comes first. Adding more -v flags will provide you with reams of debugging information.

  • --blocks=NUM: The number of 1MB packet blocks used by AF_PACKET to store packets in memory, per thread. This flag basically allows you to control how much RAM the stenotype binary uses: blocks * threads * 1MB. More blocks will allow a thread to handle traffic spikes: if you have 2048 blocks (the default), then a thread can hold 2GB of traffic in memory while waiting for it to hit disk. If you have slow links and you want to decrease memory usage, you can probably decrease this a LOT. :)

  • --fanout_type=NUM: This sets the AF_PACKET fanout type to the passed-in value. See AF_PACKET documentation for details on options here. The default should probably be fine.

  • --filter=HEX: Allows users to specify a BPF filter for packet capture... only packets which match this filter will be written by stenotype. This is NOT a human-readable BPF filter... it's a hex-encoded compiled filter. Use the supplied compile_bpf.sh script to generate this encoding from a human-readable filter.

  • --seccomp=none|trace|kill: We use seccomp to sandbox stenotype, but we've found that this can be fragile as we switch between different machine configurations. Some VMs appear to freeze while trying to set up seccomp sandboxes: for those environments, you can pass --seccomp=none in (note that this will turn off some sandboxing). If you're trying to debug a stenotype failure you think is caused by overzealous sandboxing, you can pass in --seccomp=trace, then run stenotype with strace to figure out why things are misbehaving.

  • --preallocate_file_mb=NUM: Certain file systems handle writes faster if the file has already been allocated to its eventual size. If you set this flag to 4096, then stenotype will preallocate each new packet file to this size while opening it. The file will be truncated to its actual size when closed. This should not be necessary unless you're really trying to eak out some extra speed on a file system that supports extents.

    NOTE: If you are using xfs on 4.1+ kernel, setting "--preallocate_file_mb" becomes crucial to performance. This is in that when appending to the end of file (i.e. EOF updates), kernel will serialize all operations. Please refer to commit (b9d5984 xfs: DIO write completion size updates race).

There's a number of other flags that stenotype supports, but most of them are for debugging purposes.