Git server memory requirements

I'm trying to get an idea how much memory is required for a git server that is hosting linux kernel repos.

What we're seeing is that git uses around 1GB of RAM on the server when a user does a clone of the Linux kernel source over ssh. Does this seem about right? Is this amount fixed, or arbitrary (trade-off memory for speed). It seems subsequent concurrent clones use less memory.

Is there any practical way to reduce the memory usage? We're running into occasional issues if there are multiple clones at once.

Is setting gc.auto=0 a good idea for large kernel repos? The idea is we can repack manually or in a cron on weekends rather than during user operations. However, manually running a git gc seems to use about as much memory as a user clone.

It may be that our EC2 small instance (2.5GB) is not up to the task, but would like to understand options (we can easily trade off some speed for less memory if we can) before upgrading.

Thanks, Cliff

On Fri, Mar 21, 2014 at 01:38:01PM -0400,  Cliff Brake wrote:

Speaking for Gentoo here, as we're working on our large repo migration, and this was a concern originally. So far it's best mitigated by the pack-bitmap patches in the next branch.

The initial clones become significantly faster and much less resource intensive.

I'm hoping that the packed bitmap series lands soon in the master branch and a formal release (a big +1 from me for it).