sayounara / ermia

CC for modern main-memory OLTP systems

ERMIA

Fast and Robust OLTP using Epoch-based Resource Management and Indirection Array

For more information, you may refer to our SIGMOD'16 paper (https://github.com/ermia-db/ermia/raw/master/ermia.pdf)

Environment configurations

• Software dependencies: libnuma. Install from your favorite package manager. ERMIA uses mmap with MAP_HUGETLB to allocate huge pages. MAP_HUGETLB is available after Linux 2.6.32.
• Make sure you have enough huge pages. Almost all memory allocations come from the space carved out here. Assuming 2MB pages, the command below will allocate 40GB of memory:

sudo sh -c 'echo [x pages] > /proc/sys/vm/nr_hugepages'

This limits the maximum for --node-memory-gb to 10 for a 4-socket machine (see below).

• mlock limits. Add the following to /etc/security/limits.conf (replace "[user]" with your login):

[user] soft memlock unlimited
[user] hard memlock unlimited

Re-login to apply.

Choosing a CC scheme

ERMIA supports Read Committed (RC), Snapshot Isolation (SI), Serializable Snapshot Isolation (SSI), and Serial Safety Net (SSN) with SI/RC. Switches are defined in src/macros.h:

• #define USE_PARALLEL_SSI to use serializable snapshot isolation (SSI);
• #define USE_PARALLEL_SSN to use the serial safety net (SSN);
  • SSN may be used in conjunction with RC (USE_READ_COMMITTED) or SI (default);
  • # define DO_EARLY_SSN_CHECKS enables SSN window exclusion tests during normal reads and writes.

USE_PARALLEL_SSI and USE_PARALLEL_SSN are mutually exclusive.

Giving -D[SCHEME] to $make also works. SCHEME can be RC, SI, RC_SSN, SI_SSN, or SSI.

Adjust maximum concurrent workers

By default we support up to 256 cores. The limit can be adjusted by setting MAX_THREADS defined under sysconf in dbcore/sm-config.h. MAX_THREADS must be a multiple of 64.

Build it

---

Use src/build.sh to compile ERMIA. For performance runs, $ build.sh, $ build.sh 1 for debugging.

Run it

$run.sh \
       [executable] \
       [benchmark] \
       [num-threads] \
       [duration (seconds)] \
       "[other system-wide runtime options]" \
       "[other benchmark-specific runtime options]"`

Note the quotation marks for additional options.

System-wide runtime options

--node-memory-gb: how many GBs of memory to allocate per socket.

--null-log-device: flush log buffer to /dev/null. With more than 30 threads, log flush (even to tmpfs) can easily become a bottleneck because of a mutex in the kernel held during the flush. This option does not disable logging, but it voids the ability to recover.

--tmpfs-dir: location of the log buffer's mmap file. Default: /tmpfs/.

--enable-gc: turn on garbage collection. Currently there is only one GC thread.

--enable-chkpt: enable checkpointing.

--warm-up: strategy to load versions upon recovery. Candidates are:

• eager: load all latest versions during recovery, so the database is fully in-memory when it starts to process new transactions;
• lazy: start a thread to load versions in the background after recovery, so the database is partially in-memory when it starts to process new transactions.
• none: load versions on-demand upon access.

SSI and SSN specific:

--safesnap: enable safe snapshot for read-only transactions.

SSN-specific:

--ssn-read-opt-threshold: versions that are read by a read-mostly transaction and older than this value are considered "old" and will not be tracked; setting it to 0 will skip all read tracking for read-mostly transactions (TXN_FLAG_READ_MOSTLY).

SSI-specific: --ssi-read-only-opt: enable P&G style read-only optimization for SSI.