Hacker News Reader

Comment by sumanthvepa | Hacker News Reader

The big problem with Linux at the time was large files, and memory addressing limitations of 32bit architectures. x86_64 was still too new and only introduced in 1999. There were other issues as well, that in the intervening 20 years have been more than addressed. Today's Linux is of course far superior to any OS that existed at the time.

View on HN

All Unix-like systems had exactly the same issues with large file support on 32b architectures. The solution was standardized in 1996 and adopted by most relevant OSes relatively quickly. Another issue is that support for large files has to be opt-in for user space code and this opt-in has to be same for all modules that share anything whose ABI depends on sizeof(off_t).

Server software that needs larger address space than what is conceivable size of main memory became somewhat common only after mass adoption of amd64. In fact, amd64 is first 64bit architecture where it is common to have 64-bit only userspace, essentially all "64b" unix vendors in 90's shipped systems with most of userspace built in 32b mode (IIRC this includes even technically 64b-only OSF/1 AXP^W^WTru64, which while running on CPU that really does not have 32b mode had 32b ABI in the style of linux's x32 ABI).

> shipped systems with most of userspace built in 32b mode

Gah, so that's where ARM's THUMB instruction set got its inspiration from.

All of the BSDs have 64bit off_t, they had to break the ABI for the BSD lawsuit settlement anyway and used that opportunity to make this change.

The first AMD64 processor was the Opteron released in 2003, 4 years after the announcement.

OP might have been thinking of Itanium (ia-64), a different 64 bit PC architecture from Intel. It was shockingly expensive

There was certainly development on ia64 in the kernel in 1999 under Project Trillian [1][2], although the first chip wasn't released commercially until 2001

Alternatively they may have been thinking about PAE, which was released in 1999 and allowed linux kernels to address upto 64GB on a 32 bit processor [2]

[1] https://lkml.org/lkml/1999/3/9/48

[2] https://en.wikipedia.org/wiki/Project_Trillian

[2] https://en.wikipedia.org/wiki/Physical_Address_Extension#His...

PAE was also used for Microsoft Windows 2000 Server Datacenter Edition. Server machines at the time capable of holding 64GB RAM were beasts, IIRC on the order of hundreds of thousands of dollars. A desktop machine with 512MB memory was considered way more than you needed. I used a Toshiba Tecra 9000 laptop with 256GB for mostly Java and C++ development and never felt like I was anywhere near maxing it out.

PAE might not have met Amazon's needs for holding their catalog or whatever. If they were trying to directly address a data structure that was larger than 4GB, they would have needed some sort of trickery as PAE was usually implemented kernel side to provide separate 4GB address spaces to individual processes with more than 4GB total physical backing.

> Toshiba Tecra 9000 laptop with 256GB

MB I assume :D

This was before OSes had been replaced by the web stack, which seems to bleed memory.

Linux ran on other 64-bit architectures before that. WP says "1995: First 64-bit Linux distribution for the Alpha architecture is released" in https://en.wikipedia.org/wiki/64-bit_computing#64-bit_operat...

X86_64 wasn't widely available until around 2005 which is also when Java added support for it

Reply on news.ycombinator.com