[Linux-cluster] Cluster vs Distributed? & MySQL Cluster?

[David Brieck Jr.]

On 10/25/06, Michael Will <mwill at penguincomputing.com> wrote:
> Are the actual data files shared in this setup between the active mysql
> daemons?
>
> Last time I looked into this it seemed that with shared-nothing model each
> mysql daemon would have to keep it's own copy of the data and updates
> would be propagated from active to passive daemons (master-slave model)
> or between active daemons (ndb in-ram database model)
>
> Are the mysql daemons running on the GFS I/O nodes that have access to
> shared
> storage via SAN or iSCSI and coordinate locking through GFS
> infrastructure, or are
> the mysql daemons running on client nodes that use GFS to remotely
> access storage
> that is provided by other  GFS I/O nodes that in turn have access to shared
> storage via SAN or iSCSI?
>
> Michael
>

We're using GNBD for the nodes to connect to the storage. We don't
have the fastest storage setup right now, but I'm hopeful that if
everything works well we'll be purchasing a faster storage setup.

As far as MySQL using GFS (excluding anything with active-active) and
using DLM to do locks, here are some comparisons:

Benchmark on GFS

Benchmark DBD suite: 2.15
Date of test:        2006-10-26  9:49:43
Running tests on:    Linux 2.6.9-42.0.2.ELhugemem i686
Arguments:           --small-test --tcpip --fast --fast-insert --lock-tables
Comments:
Limits from:
Server version:      MySQL 4.1.20/
Optimization:        None
Hardware:

alter-table: Total time: 94 wallclock secs ( 0.02 usr  0.01 sys +
0.00 cusr  0.00 csys =  0.03 CPU)
big-tables: Total time:  4 wallclock secs ( 0.13 usr  0.14 sys +  0.00
cusr  0.00 csys =  0.27 CPU)
connect: Total time:  5 wallclock secs ( 0.38 usr  0.53 sys +  0.00
cusr  0.00 csys =  0.91 CPU)
create: Total time:  8 wallclock secs ( 0.02 usr  0.01 sys +  0.00
cusr  0.00 csys =  0.03 CPU)
insert: Total time: 17 wallclock secs ( 2.19 usr  1.99 sys +  0.00
cusr  0.00 csys =  4.18 CPU)
select: Total time: 13 wallclock secs ( 2.36 usr  1.03 sys +  0.00
cusr  0.00 csys =  3.39 CPU)

Benchmark on Local

alter-table: Total time: 70 wallclock secs ( 0.02 usr  0.00 sys +
0.00 cusr  0.00 csys =  0.02 CPU)
big-tables: Total time:  2 wallclock secs ( 0.11 usr  0.14 sys +  0.00
cusr  0.00 csys =  0.25 CPU)
connect: Total time:  4 wallclock secs ( 0.37 usr  0.55 sys +  0.00
cusr  0.00 csys =  0.92 CPU)
create: Total time:  1 wallclock secs ( 0.01 usr  0.00 sys +  0.00
cusr  0.00 csys =  0.01 CPU)
insert: Total time: 13 wallclock secs ( 2.27 usr  1.95 sys +  0.00
cusr  0.00 csys =  4.22 CPU)
select: Total time: 12 wallclock secs ( 2.21 usr  0.97 sys +  0.00
cusr  0.00 csys =  3.18 CPU)

It's pretty darn close and I'm willing to take a small performance hit.

Here's some relevant info: local storage is RAID5 and GFS is RAID10
and shared using CLVM, multipath, and GNBD. So the speed of the test
locally would probably be faster if it were either RAID1 or 10, not 5.