Models of Golgi apparatus biogenesis and maintenance are focused on two possibilities: one is self‐assembly from the endoplasmic reticulum, and the other is nucleation by a stable template. Here, we asked in three different experimental situations whether assembly of the Golgi apparatus might be dynamically nucleated. During microtubule depolymerization, the integral membrane protein p27 and the peripheral Golgi protein GM130, appeared in newly formed, scattered Golgi elements before three different Golgi apparatus cisternal enzymes, whereas GRASP55, a medial peripheral Golgi protein, showed, if anything, a tendency to accumulate in scattered Golgi elements later than a cisternal enzyme. During Golgi formation after brefeldin A washout, endoplasmic reticulum exit of Golgi resident enzymes could be completely separated from that of p27 and GM130. p27 and GM130 accumulation was onto newly organized perinuclear structures, not brefeldin A remnants, and preceded that of a cisternal enzyme. Reassembly was completely sensitive to guanosine 5′‐diphosphate‐restricted Sar1p. When cells were microinjected with Sar1pWT DNA to reverse a guanosine 5′‐diphosphate‐restricted Sar1p endoplasmic reticulum‐exit block phenotype, GM130 and p27 collected perinuclearly with little to no exit of a cisternal enzyme from the endoplasmic reticulum. The overall data strongly indicate that the assembly of the Golgi apparatus can be nucleated dynamically by GM130/p27 associated structures. We define dynamic nucleation as the first step in a staged organelle assembly process in which new component association forms a microscopically visible structure onto which other components add later, e.g. Golgi cisternae.

Access to supercomputers is generally limited to dedicated terminals or a workstation. With the advent of local area networks and resource sharing, many distributed applications have been developed. This paper presents two models for integrating workstations and a vector processor on a local area network. A compute server which consists of a dedicated workstation with a link to the vector processor, is used to provide the vector processor service. Implementations are developed for each model and are tested using alternative transport protocols, namely TCP and UDP. Experimental results indicate that the vector processor service suffers negligible degradation due to the network communication overhead in comparision to standalone operation. The model based on remote execution is slightly faster than the model based on remote procedure call due to less communication overhead.